---

---

CASE STUDY - 1 :: Healthcare Provider Fraudulent Detection

---

---

1. This notebook contains the extensive data analysis of { BENE + IP + OP + Frauds Providers Labels } performed on the publicly available dataset at Kaggle with the intent:

• To gain better understanding of the dataset from different levels.
• To explore the dataset in order to uncover the potential hidden patterns.
• And, to find the answers of various WHYs.

2. I have also added some new features with the intent to bring some more business aspects in the dataset. Also, performed their impact analysis on the potential frauds.

Kindly checkout the below Deck for better understanding the BUSINESS oriented insights about this problem:

• Deck : Detailed Explanation

Kindly checkout the below Doc for TECHNICAL Design description about this problem:

• Technical Document

Notebook Contents

CASE STUDY - 1 :: Healthcare Provider Fraudulent Detection

Notebook Contents

Downloading_Train_Data_Files

TRAIN set files

Importing_Libraries

Importing_Dataset

Exploring_Target_Labels_Data

Adding the Admitted or Not Admitted indicator in IP and OP Dataset

Merging the Datasets

Merging the IP_OP Dataset with BENE Data

Merging the IP_OP_BENE Dataset with PROVIDER level Tgt Labels Data

Entire Dataset

ASSUMPTION :: One provider may have been involved in more than one claim. So, does all the claims filed by a potentially fraud provider are all frauds?

Therefore, it is a big assumption to make that all the claims filed by a potentially fraud provider are fraudulent.

Feature Engineering + Impact Analysis

Adding New Feature :: Is_Alive?

Adding New Feature :: Claim_Duration

Adding New Feature :: Admitted_Duration

Adding New Feature :: Bene_Age

Does InscClaimAmtReimbursed influences Potentially Fraud?

Does IPAnnualReimbursementAmt influences Potentially Fraud?

Why do we have IP Annual Re-Imb Amount as 0 for Admitted Patients?

Does OPAnnualReimbursementAmt influences Potentially Fraud?

Why do we have OP Annual Re-Imb Amount as 0 for Admitted Patients?

Adding New Feature :: Total Number of false claims filed by a Provider

Adding New Feature :: Total Number of claims or cases seen by Attending Physician

Adding New Feature :: Total Number of claims or cases seen by Opearting Physician

Adding New Feature :: Total Number of claims or cases seen by Other Physician

Adding Combined Feature :: Att_Opr_Oth_Phy_Tot_Claims

Adding 3 New Features :: Prv_Tot_Att_Phy, Prv_Tot_Opr_Phy and Prv_Tot_Oth_Phy

Adding Combined Feature :: Prv_Tot_Att_Opr_Oth_Phys

Adding New Feature :: Total Unique Claim Admit Codes used by a PROVIDER

Adding New Feature :: Total Unique Number of Diagnosis Group Codes used by a PROVIDER

Adding New Feature :: Total unique Date of Birth years of beneficiaries provided by a Provider

Adding New Feature :: Sum of patients age treated by a Provider

Adding New Feature :: Sum of Insc Claim Re-Imb Amount for a Provider

Adding New Feature :: Total number of RKD Patients seen by a Provider

Exploratory Data Analysis

Q1. Which are the Top-25 Providers with maximum number of fraudulent cases?

Q2. Which are the Top-25 Providers with maximum number of non-fraudulent cases?

Q3. Which are the Top-25 Attending Physicians with maximum number of fraudulent cases?

Q4. Which are the Top-25 Attenting Physicians with maximum number of non-fraudulent cases?

Q5. Which are the Top-25 Operating Physicians with maximum number of fraudulent cases?

Q6. Which are the Top-25 Operating Physicians with maximum number of non-fraudulent cases?

Q7. Which are the Top-25 Other Physicians with maximum number of fraudulent cases?

Q8. Which are the Top-25 Other Physicians with maximum number of non-fraudulent cases?

Q9. Which are the Top-25 ClmAdmitDiagnosisCode with maximum number of fraudulent cases?

Q10. Which are the Top-25 ClmAdmitDiagnosisCode with maximum number of non-fraudulent cases?

Q11. Which are the Top-25 DiagnosisGroupCode with maximum number of fraudulent cases?

Q12. Which are the Top-25 DiagnosisGroupCode with maximum number of non-fraudulent cases?

Q13. Does Age_groups have any relationship with maximum number of fraudulent cases?

Q14. Does Age_groups have any relationship with maximum number of non-fraudulent cases?

Q15. Which are the Top-25 States with maximum number of fraudulent cases?

Q16. What are the Top-25 States with maximum number of non-fraudulent cases?

Q17. Which are the Top-25 Country with maximum number of fraudulent cases?

Q18. What are the Top-25 Country with maximum number of non-fraudulent cases?

Q19. Does various Human Races have any relationship with maximum number of fraudulent cases?

Q20. Does various Human Races have any relationship with maximum number of non-fraudulent cases?

Downloading_Train_Data_Files

## TRAIN set files

!gdown 12zSQN2FOxmuXFhz2xzPNussPisEfVP5w
!gdown 13XyBakfHiG-BNQPrYFXAHlsOcfICOTpx
!gdown 1dLxl4vkykPcm4Zj0abYR0Ohr7STQHg-1
!gdown 1rFER-7VuYb7GfCYeJrfxPidgK0lwqw3R

Importing_Libraries

import os
import sys
import math
import scipy as scipy
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns

%matplotlib inline

pd.set_option('display.max_columns',80)
label_font_dict = {'family':'sans-serif','size':13.5,'color':'brown','style':'italic'}
title_font_dict = {'family':'sans-serif','size':16.5,'color':'Blue','style':'italic'}

Importing_Dataset

train_bene_df = pd.read_csv("Dataset/Train/Train_Beneficiarydata-1542865627584.csv")
train_ip_df = pd.read_csv("Dataset/Train/Train_Inpatientdata-1542865627584.csv")
train_op_df = pd.read_csv("Dataset/Train/Train_Outpatientdata-1542865627584.csv")
train_tgt_lbls_df = pd.read_csv("Dataset/Train/Train-1542865627584.csv")

Exploring_Target_Labels_Data

train_tgt_lbls_df.head()

	Provider	PotentialFraud
0	PRV51001	No
1	PRV51003	Yes
2	PRV51004	No
3	PRV51005	Yes
4	PRV51007	No

• Check the Fraud and Non-Fraud Counts

print("### The unique number of providers are {}. ###".format(train_tgt_lbls_df.shape[0]))

### The unique number of providers are 5410. ###

with plt.style.context('seaborn-poster'):
    fig = train_tgt_lbls_df["PotentialFraud"].value_counts().plot(kind='bar', color=['green','orange'])
    # Using the "patches" function we will get the location of the rectangle bars from the graph.
    ## Then by using those location(width & height) values we will add the annotations
    for p in fig.patches:
        width = p.get_width()
        height = p.get_height()
        x, y = p.get_xy()
        fig.annotate(f'{str(round((height*100)/train_tgt_lbls_df.shape[0],2))+"%"}', (x + width/2, y + height*1.015), ha='center', fontsize=13.5)
    # Providing the labels and title to the graph
    plt.xlabel("Provider Fraud or Not?", fontdict=label_font_dict)
    plt.ylabel("Number or % share of providers\n", fontdict=label_font_dict)
    plt.yticks(np.arange(0,5100,500))
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.minorticks_on()
    plt.title("Distribution of Fraud & Non-fraud providers\n", fontdict=title_font_dict)
    plt.plot();

OBSERVATION

• From the above plot, we can say that 90% of the providers are not frausters and only 9% of them are involved in frauds.

Adding the Admitted or Not Admitted indicator in IP and OP Dataset

• Adding in IP Dataset

train_ip_df["Admitted?"] = 1

train_ip_df.head()

	BeneID	ClaimID	ClaimStartDt	ClaimEndDt	Provider	InscClaimAmtReimbursed	AttendingPhysician	OperatingPhysician	OtherPhysician	AdmissionDt	ClmAdmitDiagnosisCode	DeductibleAmtPaid	DischargeDt	DiagnosisGroupCode	ClmDiagnosisCode_1	ClmDiagnosisCode_2	ClmDiagnosisCode_3	ClmDiagnosisCode_4	ClmDiagnosisCode_5	ClmDiagnosisCode_6	ClmDiagnosisCode_7	ClmDiagnosisCode_8	ClmDiagnosisCode_9	ClmDiagnosisCode_10	ClmProcedureCode_1	ClmProcedureCode_2	ClmProcedureCode_3	ClmProcedureCode_4	ClmProcedureCode_5	ClmProcedureCode_6	Admitted?
0	BENE11001	CLM46614	2009-04-12	2009-04-18	PRV55912	26000	PHY390922	NaN	NaN	2009-04-12	7866	1068.0	2009-04-18	201	1970	4019	5853	7843	2768	71590	2724	19889	5849	NaN	NaN	NaN	NaN	NaN	NaN	NaN	1
1	BENE11001	CLM66048	2009-08-31	2009-09-02	PRV55907	5000	PHY318495	PHY318495	NaN	2009-08-31	6186	1068.0	2009-09-02	750	6186	2948	56400	NaN	NaN	NaN	NaN	NaN	NaN	NaN	7092.0	NaN	NaN	NaN	NaN	NaN	1
2	BENE11001	CLM68358	2009-09-17	2009-09-20	PRV56046	5000	PHY372395	NaN	PHY324689	2009-09-17	29590	1068.0	2009-09-20	883	29623	30390	71690	34590	V1581	32723	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	1
3	BENE11011	CLM38412	2009-02-14	2009-02-22	PRV52405	5000	PHY369659	PHY392961	PHY349768	2009-02-14	431	1068.0	2009-02-22	067	43491	2762	7843	32723	V1041	4254	25062	40390	4019	NaN	331.0	NaN	NaN	NaN	NaN	NaN	1
4	BENE11014	CLM63689	2009-08-13	2009-08-30	PRV56614	10000	PHY379376	PHY398258	NaN	2009-08-13	78321	1068.0	2009-08-30	975	042	3051	34400	5856	42732	486	5119	29620	20300	NaN	3893.0	NaN	NaN	NaN	NaN	NaN	1

• Adding in OP Dataset

train_op_df["Admitted?"] = 0

train_op_df.head()

	BeneID	ClaimID	ClaimStartDt	ClaimEndDt	Provider	InscClaimAmtReimbursed	AttendingPhysician	OperatingPhysician	OtherPhysician	ClmDiagnosisCode_1	ClmDiagnosisCode_2	ClmDiagnosisCode_3	ClmDiagnosisCode_4	ClmDiagnosisCode_5	ClmDiagnosisCode_6	ClmDiagnosisCode_7	ClmDiagnosisCode_8	ClmDiagnosisCode_9	ClmDiagnosisCode_10	ClmProcedureCode_1	ClmProcedureCode_2	ClmProcedureCode_3	ClmProcedureCode_4	ClmProcedureCode_5	ClmProcedureCode_6	DeductibleAmtPaid	ClmAdmitDiagnosisCode	Admitted?
0	BENE11002	CLM624349	2009-10-11	2009-10-11	PRV56011	30	PHY326117	NaN	NaN	78943	V5866	V1272	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	0	56409	0
1	BENE11003	CLM189947	2009-02-12	2009-02-12	PRV57610	80	PHY362868	NaN	NaN	6115	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	0	79380	0
2	BENE11003	CLM438021	2009-06-27	2009-06-27	PRV57595	10	PHY328821	NaN	NaN	2723	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	0	NaN	0
3	BENE11004	CLM121801	2009-01-06	2009-01-06	PRV56011	40	PHY334319	NaN	NaN	71988	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	0	NaN	0
4	BENE11004	CLM150998	2009-01-22	2009-01-22	PRV56011	200	PHY403831	NaN	NaN	82382	30000	72887	4280	7197	V4577	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	0	71947	0

Merging the Datasets

# Commom columns must be 28
common_cols = [col for col in train_ip_df.columns if col in train_op_df.columns]
len(common_cols)

28

# Merging the IP and OP dataset on the basis of common columns
train_ip_op_df = pd.merge(left=train_ip_df, right=train_op_df, left_on=common_cols, right_on=common_cols, how="outer")
train_ip_op_df.shape

(558211, 31)

train_ip_op_df.head()

	BeneID	ClaimID	ClaimStartDt	ClaimEndDt	Provider	InscClaimAmtReimbursed	AttendingPhysician	OperatingPhysician	OtherPhysician	AdmissionDt	ClmAdmitDiagnosisCode	DeductibleAmtPaid	DischargeDt	DiagnosisGroupCode	ClmDiagnosisCode_1	ClmDiagnosisCode_2	ClmDiagnosisCode_3	ClmDiagnosisCode_4	ClmDiagnosisCode_5	ClmDiagnosisCode_6	ClmDiagnosisCode_7	ClmDiagnosisCode_8	ClmDiagnosisCode_9	ClmDiagnosisCode_10	ClmProcedureCode_1	ClmProcedureCode_2	ClmProcedureCode_3	ClmProcedureCode_4	ClmProcedureCode_5	ClmProcedureCode_6	Admitted?
0	BENE11001	CLM46614	2009-04-12	2009-04-18	PRV55912	26000	PHY390922	NaN	NaN	2009-04-12	7866	1068.0	2009-04-18	201	1970	4019	5853	7843	2768	71590	2724	19889	5849	NaN	NaN	NaN	NaN	NaN	NaN	NaN	1
1	BENE11001	CLM66048	2009-08-31	2009-09-02	PRV55907	5000	PHY318495	PHY318495	NaN	2009-08-31	6186	1068.0	2009-09-02	750	6186	2948	56400	NaN	NaN	NaN	NaN	NaN	NaN	NaN	7092.0	NaN	NaN	NaN	NaN	NaN	1
2	BENE11001	CLM68358	2009-09-17	2009-09-20	PRV56046	5000	PHY372395	NaN	PHY324689	2009-09-17	29590	1068.0	2009-09-20	883	29623	30390	71690	34590	V1581	32723	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	1
3	BENE11011	CLM38412	2009-02-14	2009-02-22	PRV52405	5000	PHY369659	PHY392961	PHY349768	2009-02-14	431	1068.0	2009-02-22	067	43491	2762	7843	32723	V1041	4254	25062	40390	4019	NaN	331.0	NaN	NaN	NaN	NaN	NaN	1
4	BENE11014	CLM63689	2009-08-13	2009-08-30	PRV56614	10000	PHY379376	PHY398258	NaN	2009-08-13	78321	1068.0	2009-08-30	975	042	3051	34400	5856	42732	486	5119	29620	20300	NaN	3893.0	NaN	NaN	NaN	NaN	NaN	1

Merging the IP_OP Dataset with BENE Data

# Joining the IP_OP dataset with the BENE data
train_ip_op_bene_df = pd.merge(left=train_ip_op_df, right=train_bene_df, left_on='BeneID', right_on='BeneID',how='inner')
train_ip_op_bene_df.shape

(558211, 55)

Merging the IP_OP_BENE Dataset with PROVIDER level Tgt Labels Data

# Joining the IP_OP_BENE dataset with the Tgt Label Provider Data
train_iobp_df = pd.merge(left=train_ip_op_bene_df, right=train_tgt_lbls_df, left_on='Provider', right_on='Provider',how='inner')
train_iobp_df.shape

(558211, 56)

Entire Dataset

train_iobp_df.shape

(558211, 56)

# Unique Providers
train_iobp_df["Provider"].nunique()

5410

# Unique Claim numbers
train_iobp_df["ClaimID"].nunique()

558211

ASSUMPTION :: One provider may have been involved in more than one claim. So, does all the claims filed by a potentially fraud provider are all frauds?

- This cannot holds True for all the providers because if one provider has filed say 50 claims then we can't say that all the claims for that provider are fraudulent. 
    - There may exists a pattern that out of 50 claims a provider files 1 or 2 fraudulent claims. 

Therefore, it is a big assumption to make that all the claims filed by a potentially fraud provider are fraudulent.

prvs_claims_df = pd.DataFrame(train_iobp_df.groupby(['Provider'])['ClaimID'].count()).reset_index()
prvs_claims_tgt_lbls_df = pd.merge(left=prvs_claims_df, right=train_tgt_lbls_df, on='Provider', how='inner')
prvs_claims_tgt_lbls_df

	Provider	ClaimID	PotentialFraud
0	PRV51001	25	No
1	PRV51003	132	Yes
2	PRV51004	149	No
3	PRV51005	1165	Yes
4	PRV51007	72	No
...	...	...	...
5405	PRV57759	28	No
5406	PRV57760	22	No
5407	PRV57761	82	No
5408	PRV57762	1	No
5409	PRV57763	118	No

5410 rows × 3 columns

OBSERVATION

• As shown in the above table, PRV51005 has filed 1165 claims so after joining the datasets all of these will be marked as Fraud.

• Fraud Count at Claims level

print(pd.DataFrame(train_iobp_df['PotentialFraud'].value_counts()), "\n")

with plt.style.context('seaborn-poster'):
    fig = train_iobp_df['PotentialFraud'].value_counts().plot(kind='bar', color=['green','orange'])
    # Using the "patches" function we will get the location of the rectangle bars from the graph.
    ## Then by using those location(width & height) values we will add the annotations
    for p in fig.patches:
        width = p.get_width()
        height = p.get_height()
        x, y = p.get_xy()
        fig.annotate(f'{str(round((height*100)/train_iobp_df.shape[0],2))+"%"}', (x + width/2, y + height*1.015), ha='center', fontsize=13.5)
    # Providing the labels and title to the graph
    plt.xlabel("Fraud or Not?", fontdict=label_font_dict)
    plt.ylabel("Number (or %) of claims\n", fontdict=label_font_dict)
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.minorticks_on()
    plt.title("Distribution of Fraud & Non-fraud claims\n", fontdict=title_font_dict)
    plt.plot();

     PotentialFraud
No           345415
Yes          212796 

OBSERVATION

• The above plot shows us that, 62% of claims are Non-Fraud and 32% of them are Fraudulent.
  • By looking at the percentages we may say that there is a class-imbalance problem but after looking at the number of records it doesn't seem to be a severe class-imbalance problem.
    • So, I'll try some class balancing techniques only after training a baseline model w/o any synthetic or class weighting techniques.

Feature Engineering + Impact Analysis

Let's create some features

Adding New Feature :: Is_Alive?

- Is Alive? = No if DOD is NaN else Yes

train_iobp_df['DOB'] = pd.to_datetime(train_iobp_df['DOB'], format="%Y-%m-%d")
train_iobp_df['DOD'] = pd.to_datetime(train_iobp_df['DOD'], format="%Y-%m-%d")

train_iobp_df['Is_Alive?'] = train_iobp_df['DOD'].apply(lambda val: 'No' if val != val else 'Yes')

train_iobp_df['Is_Alive?'].value_counts()

No     554080
Yes      4131
Name: Is_Alive?, dtype: int64

Adding New Feature :: Claim_Duration

- Claim Duration = Claim End Date - Claim Start Date

train_iobp_df['ClaimStartDt'] = pd.to_datetime(train_iobp_df['ClaimStartDt'], format="%Y-%m-%d")
train_iobp_df['ClaimEndDt'] = pd.to_datetime(train_iobp_df['ClaimEndDt'], format="%Y-%m-%d")

train_iobp_df['Claim_Duration'] = (train_iobp_df['ClaimEndDt'] - train_iobp_df['ClaimStartDt']).dt.days

with plt.style.context('seaborn'):
    fig = sns.boxenplot(data=train_iobp_df, x='PotentialFraud',y='Claim_Duration', palette='dark')
    # Providing the labels and title to the graph
    plt.xlabel("Potentially Fraud?", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Claim Duration (in days)\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Claim Duration for Potentially Fraud & Non-Fraud Providers\n", fontdict=title_font_dict)
    plt.plot();

OBSERVATION

• The above plot clearly shows us that there is no difference in the distribution of Claim Duration for Potentially Fraud and Non-Fraud Providers.
  • Therefore, we can say that Claim Duration alone might not be useful in segregating the Fraud cases.

• Relationship b/w Claim_Duration and Potentially Fraud for both the Genders

with plt.style.context('seaborn-poster'):
    fig = sns.boxenplot(data=train_iobp_df, x='PotentialFraud',y='Claim_Duration', hue='Gender', palette='prism')      
    # Providing the labels and title to the graph
    plt.xlabel("Potentially Fraud?", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Claim Duration (in days)\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Claim Duration of both the genders for Potentially Fraud & Non-Fraud Providers\n", fontdict=title_font_dict)
    plt.plot();

OBSERVATION

• The above plot clearly shows us that there is no difference in the distribution of Claim Duration of males and females for Potentially Fraud and Non-Fraud Providers.
  • Therefore, we can say that Claim Duration might not be useful in segregating the Fraud cases.

• Relationship b/w Claim_Duration and Potentially Fraud for Is_Alive?

with plt.style.context('seaborn-poster'):
    fig = sns.boxenplot(data=train_iobp_df, x='PotentialFraud',y='Claim_Duration', hue='Is_Alive?', palette='prism')      
    # Providing the labels and title to the graph
    plt.xlabel("Potentially Fraud?", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Claim Duration (in days)\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Claim Duration patient life status for Potentially Fraud & Non-Fraud Providers\n", fontdict=title_font_dict)
    plt.plot();

OBSERVATION

• The above plot clearly shows us that there is no difference in the distribution of Claim Duration on patient life status for Potentially Fraud and Non-Fraud Providers.
  • Therefore, we can say that Claim Duration might not be useful in segregating the Fraud cases.

• Relationship b/w Claim_Duration and Potentially Fraud for all Human Races

with plt.style.context('seaborn-poster'):
    plt.figure(figsize=(16,8))
    fig = sns.boxenplot(data=train_iobp_df, x='PotentialFraud',y='Claim_Duration', hue='Race', palette='cubehelix')   
    # Providing the labels and title to the graph
    plt.xlabel("Potentially Fraud?", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Claim Duration (in days)\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Claim Duration of all human races for Potentially Fraud & Non-Fraud Providers\n", fontdict=title_font_dict)
    plt.legend(loc='upper center',title='Race');

OBSERVATION

• The above plot clearly shows us that there is no difference in the distribution of Claim Duration of all human races for Potentially Fraud and Non-Fraud Providers.
  • Therefore, we can say that Claim Duration might not be useful in segregating the Fraud cases.

• Relationship b/w Claim_Duration and Potentially Fraud for RenalDiseaseIndicator

with plt.style.context('seaborn-poster'):
    fig = sns.boxenplot(data=train_iobp_df, x='PotentialFraud',y='Claim_Duration', hue='RenalDiseaseIndicator', palette='copper')   
    # Providing the labels and title to the graph
    plt.xlabel("Potentially Fraud?", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Claim Duration (in days)\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Claim Duration of whether having RKD influence Potentially Fraud & Non-Fraud Providers?\n", fontdict=title_font_dict)
    plt.legend(loc="upper center", title='RKD');

OBSERVATION

• The above plot clearly shows us that there is no difference in the distribution of Claim Duration of patients with or w/o RKD for Potentially Fraud and Non-Fraud Providers.
  • Therefore, we can say that Claim Duration might not be useful in segregating the Fraud cases.

Adding New Feature :: Admitted_Duration

- Admitted Duration = Discharge Date - Admission Date

train_iobp_df['AdmissionDt'] = pd.to_datetime(train_iobp_df['AdmissionDt'], format="%Y-%m-%d")
train_iobp_df['DischargeDt'] = pd.to_datetime(train_iobp_df['DischargeDt'], format="%Y-%m-%d")

train_iobp_df['Admitted_Duration'] = (train_iobp_df['DischargeDt'] - train_iobp_df['AdmissionDt']).dt.days

with plt.style.context('seaborn'):
    fig = sns.violinplot(data=train_iobp_df, x='PotentialFraud',y='Admitted_Duration', palette='Accent_r')   
    # Providing the labels and title to the graph
    plt.xlabel("Potentially Fraud?", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Admitted Duration (in days)\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Admitted Duration for Potentially Fraud & Non-Fraud Providers\n", fontdict=title_font_dict)
    plt.plot();

OBSERVATION

• The above plot clearly shows us that there is no difference in the distribution of Admit Duration for Potentially Fraud and Non-Fraud Providers.
  • Therefore, we can say that Admit Duration alone might not be useful in segregating the Fraud cases.

• Relationship b/w Admit_Duration and Potentially Fraud for both the Genders

with plt.style.context('seaborn-poster'):
    fig = sns.violinplot(data=train_iobp_df, x='PotentialFraud',y='Admitted_Duration', hue='Gender', palette='inferno')   
    # Providing the labels and title to the graph
    plt.xlabel("Potentially Fraud?", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Admit Duration (in days)\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Admit Duration of both the genders for Potentially Fraud & Non-Fraud Providers\n", fontdict=title_font_dict)
    plt.plot();

OBSERVATION

• The above plot clearly shows us that there is no difference in the distribution of Admit Duration of males and females for Potentially Fraud and Non-Fraud Providers.
  • Therefore, we can say that Admit Duration might not be useful in segregating the Fraud cases.

• Relationship b/w Admitted_Duration and Potentially Fraud for Is_Alive?

with plt.style.context('seaborn-poster'):
    fig = sns.violinplot(data=train_iobp_df, x='PotentialFraud',y='Admitted_Duration', hue='Is_Alive?', palette='prism')      
    # Providing the labels and title to the graph
    plt.xlabel("Potentially Fraud?", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Admit Duration (in days)\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Admit Duration patient life status for Potentially Fraud & Non-Fraud Providers\n", fontdict=title_font_dict)
    plt.legend(loc="upper center", title='Is_Alive?');

OBSERVATION

• The above plot clearly shows us that there is no difference in the distribution of Admit Duration on patient life status for Potentially Fraud and Non-Fraud Providers.
  • Therefore, we can say that Admit Duration might not be useful in segregating the Fraud cases.

• Relationship b/w Admit_Duration and Potentially Fraud for all Human Races

with plt.style.context('seaborn-poster'):
    fig = sns.violinplot(data=train_iobp_df, x='PotentialFraud',y='Admitted_Duration', hue='Race', palette='plasma')   
    # Providing the labels and title to the graph
    plt.xlabel("Potentially Fraud?", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Admit Duration (in days)\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Admit Duration of all human races for Potentially Fraud & Non-Fraud Providers\n", fontdict=title_font_dict)
    plt.legend(loc="upper center", title="Race");

OBSERVATION

• The above plot clearly shows us that there is no difference in the distribution of Admit Duration of all human races for Potentially Fraud and Non-Fraud Providers.
  • Therefore, we can say that Admit Duration might not be useful in segregating the Fraud cases.

• Relationship b/w Admitted_Duration and Potentially Fraud for RenalDiseaseIndicator

with plt.style.context('seaborn-poster'):
    fig = sns.violinplot(data=train_iobp_df, x='PotentialFraud',y='Admitted_Duration', hue='RenalDiseaseIndicator',palette='magma')   
    # Providing the labels and title to the graph
    plt.xlabel("Potentially Fraud?", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Admit Duration (in days)\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Admit Duration of whether having RKD influence Potentially Fraud & Non-Fraud Providers?\n", fontdict=title_font_dict)
    plt.legend(loc="upper center", title="RKD");

OBSERVATION

• The above plot clearly shows us that there is no difference in the distribution of Admit Duration of patients with or w/o RKD for Potentially Fraud and Non-Fraud Providers.
  • Therefore, we can say that Admit Duration might not be useful in segregating the Fraud cases.

Adding New Feature :: Bene_Age

- Bene Age = DOD - DOB (if DOD is Null then replace it with MAX date in DOD)

# Filling the Null values as MAX Date of Death in the Dataset
train_iobp_df['DOD'].fillna(value=train_iobp_df['DOD'].max(), inplace=True)

train_iobp_df['Bene_Age'] = round(((train_iobp_df['DOD'] - train_iobp_df['DOB']).dt.days)/365,1)

with plt.style.context('seaborn-poster'):
    fig = sns.violinplot(data=train_iobp_df, x='PotentialFraud',y='Bene_Age', palette='Pastel2')   
    # Providing the labels and title to the graph
    plt.xlabel("Potentially Fraud?", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Beneficiary Age (in years)\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Beneficiary Age for Potentially Fraud & Non-Fraud Providers\n", fontdict=title_font_dict)
    plt.plot();

OBSERVATION

• The above plot clearly shows us that there is no difference in the distribution of Beneficiary Age for Potentially Fraud and Non-Fraud Providers.
  • Therefore, we can say that Beneficiary Age alone might not be useful in segregating the Fraud cases.

• Relationship b/w Bene_Age and Potentially Fraud for both the Genders

with plt.style.context('seaborn-poster'):
    fig = sns.violinplot(data=train_iobp_df, x='PotentialFraud',y='Bene_Age', hue='Gender', palette='inferno')   
    # Providing the labels and title to the graph
    plt.xlabel("Potentially Fraud?", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Bene_Age (in years)\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Bene_Age of both the genders for Potentially Fraud & Non-Fraud Providers\n", fontdict=title_font_dict)
    plt.plot();

OBSERVATION

• The above plot clearly shows us that there is no difference in the distribution of Bene_Age of males and females for Potentially Fraud and Non-Fraud Providers.
  • Therefore, we can say that Bene_Age might not be useful in segregating the Fraud cases.

• Relationship b/w Bene_Age and Potentially Fraud for Is_Alive?

with plt.style.context('seaborn-poster'):
    fig = sns.violinplot(data=train_iobp_df, x='PotentialFraud',y='Bene_Age', hue='Is_Alive?', palette='prism')      
    # Providing the labels and title to the graph
    plt.xlabel("Potentially Fraud?", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Bene_Age (in years)\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Bene_Age patient life status for Potentially Fraud & Non-Fraud Providers\n", fontdict=title_font_dict)
    plt.legend(loc="upper center", title='Is_Alive?');

OBSERVATION

• The above plot clearly shows us that there is no difference in the distribution of Bene_Age on patient life status for Potentially Fraud and Non-Fraud Providers.
  • Therefore, we can say that Bene_Age might not be useful in segregating the Fraud cases.

• Relationship b/w Bene_Age and Potentially Fraud for all Human Races

with plt.style.context('seaborn-poster'):
    fig = sns.violinplot(data=train_iobp_df, x='PotentialFraud',y='Bene_Age', hue='Race', palette='plasma')   
    # Providing the labels and title to the graph
    plt.xlabel("Potentially Fraud?", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Bene_Age (in years)\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Bene_Age of all human races for Potentially Fraud & Non-Fraud Providers\n", fontdict=title_font_dict)
    plt.legend(loc="upper center", title="Race");

OBSERVATION

• The above plot clearly shows us that there is no difference in the distribution of Bene_Age of all human races for Potentially Fraud and Non-Fraud Providers.
  • Therefore, we can say that Bene_Age might not be useful in segregating the Fraud cases.

• Relationship b/w Bene_Age and Potentially Fraud for RenalDiseaseIndicator

with plt.style.context('seaborn-poster'):
    fig = sns.violinplot(data=train_iobp_df, x='PotentialFraud',y='Bene_Age', hue='RenalDiseaseIndicator',palette='magma')   
    # Providing the labels and title to the graph
    plt.xlabel("Potentially Fraud?", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Bene_Age (in years)\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Bene_Age of whether having RKD influence Potentially Fraud & Non-Fraud Providers?\n", fontdict=title_font_dict)
    plt.legend(loc='lower center', title='RKD');

OBSERVATION

• The above plot clearly shows us that there is no difference in the distribution of Bene_Age of patients with or w/o RKD for Potentially Fraud and Non-Fraud Providers.
  • Therefore, we can say that Bene_Age might not be useful in segregating the Fraud cases.

Does InscClaimAmtReimbursed influences Potentially Fraud?

• Relationship b/w InscClaimAmtReimbursed and Potentially Fraud

with plt.style.context('seaborn'):
    fig = sns.boxenplot(data=train_iobp_df, x='PotentialFraud',y='InscClaimAmtReimbursed', palette='flag')   
    # Providing the labels and title to the graph
    plt.xlabel("Potentially Fraud?", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Claim Re-Imb Amount\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Claim Re-Imb Amount for Potentially Fraud & Non-Fraud Providers\n", fontdict=title_font_dict)
    plt.plot();

OBSERVATION

• The above plot clearly shows us that there is no difference in the distribution of Claim Re-Imb Amount for Potentially Fraud and Non-Fraud Providers.
  • Therefore, we can say that Claim Re-Imb Amount alone might not be useful in segregating the Fraud cases.

• Relationship b/w InscClaimAmtReimbursed and Potentially Fraud for both the Genders

with plt.style.context('seaborn-poster'):
    fig = sns.boxenplot(data=train_iobp_df, x='PotentialFraud',y='InscClaimAmtReimbursed', hue='Gender', palette='flare')   
    # Providing the labels and title to the graph
    plt.xlabel("Potentially Fraud?", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Claim Re-Imb Amount\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Claim Re-Imb Amount of both the genders for Potentially Fraud & Non-Fraud Providers\n", fontdict=title_font_dict)
    plt.plot();

OBSERVATION

• The above plot clearly shows us that there is no difference in the distribution of Claim Re-Imb Amount of males and females for Potentially Fraud and Non-Fraud Providers.
  • Therefore, we can say that Claim Re-Imb Amount might not be useful in segregating the Fraud cases.

• Relationship b/w InscClaimAmtReimbursed and Potentially Fraud for Is_Alive?

with plt.style.context('seaborn-poster'):
    fig = sns.boxenplot(data=train_iobp_df, x='PotentialFraud',y='InscClaimAmtReimbursed', hue='Is_Alive?', palette='prism')      
    # Providing the labels and title to the graph
    plt.xlabel("Potentially Fraud?", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Claim Re-Imb Amount\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Claim Re-Imb Amount patient life status for Potentially Fraud & Non-Fraud Providers\n", fontdict=title_font_dict)
    plt.legend(loc="upper center", title='Is_Alive?');

OBSERVATION

• The above plot clearly shows us that there is no difference in the distribution of Claim Re-Imb Amount on patient life status for Potentially Fraud and Non-Fraud Providers.
  • Therefore, we can say that Claim Re-Imb Amount might not be useful in segregating the Fraud cases.

• Relationship b/w InscClaimAmtReimbursed and Potentially Fraud for all Human Races

with plt.style.context('seaborn-poster'):
    fig = sns.boxenplot(data=train_iobp_df, x='PotentialFraud',y='InscClaimAmtReimbursed', hue='Race', palette='plasma')   
    # Providing the labels and title to the graph
    plt.xlabel("Potentially Fraud?", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Claim Re-Imb Amount\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Claim Re-Imb Amount of all human races for Potentially Fraud & Non-Fraud Providers\n", fontdict=title_font_dict)
    plt.legend(loc="upper center", title="Race");

OBSERVATION

• The above plot clearly shows us that there is no difference in the distribution of Claim Re-Imb Amount of all human races for Potentially Fraud and Non-Fraud Providers.
  • Therefore, we can say that Claim Re-Imb Amount might not be useful in segregating the Fraud cases.

• Relationship b/w InscClaimAmtReimbursed and Potentially Fraud for RenalDiseaseIndicator

with plt.style.context('seaborn-poster'):
    fig = sns.boxenplot(data=train_iobp_df, x='PotentialFraud',y='InscClaimAmtReimbursed', hue='RenalDiseaseIndicator',palette='magma')   
    # Providing the labels and title to the graph
    plt.xlabel("Potentially Fraud?", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Claim Re-Imb Amount\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Claim Re-Imb Amount of whether having RKD influence Potentially Fraud & Non-Fraud Providers?\n", fontdict=title_font_dict)
    plt.legend(loc='upper center', title='RKD');

OBSERVATION

• The above plot clearly shows us that there is no difference in the distribution of Claim Re-Imb Amount of patients with or w/o RKD for Potentially Fraud and Non-Fraud Providers.
  • Therefore, we can say that claim Re-Imb Amount might not be useful in segregating the Fraud cases.

Does IPAnnualReimbursementAmt influences Potentially Fraud?

• Relationship b/w IPAnnualReimbursementAmt and Potentially Fraud

with plt.style.context('seaborn-poster'):
    fig = sns.boxenplot(data=train_iobp_df[train_iobp_df['Admitted?'] == 1], y='PotentialFraud',x='IPAnnualReimbursementAmt', 
                         palette='dark', orient='h')   
    # Providing the labels and title to the graph
    plt.ylabel("Potentially Fraud?\n", fontdict=label_font_dict)
    plt.yticks(rotation=90, fontsize=12)
    plt.xlabel("\nAnnual IP Re-Imb Amount", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.title("Annual IP Re-Imb Amount for Potentially Fraud & Non-Fraud Providers\n", fontdict=title_font_dict)
    plt.plot();

OBSERVATION

• The above plot clearly shows us that there is no difference in the distribution of Annual IP Re-Imb Amount for Potentially Fraud and Non-Fraud Providers.
  • And, it looks like a Pareto Distribution.

Why do we have IP Annual Re-Imb Amount as 0 for Admitted Patients?

print(pd.DataFrame(train_iobp_df[(train_iobp_df['IPAnnualReimbursementAmt'] == 0)]['Admitted?'].value_counts()))

   Admitted?
0     371263
1        413

• So, we have 413 claims where Patients were admitted to the hospital but the allocated IP Annual Re-Imb Amt is 0.

print(pd.DataFrame(train_iobp_df[(train_iobp_df['IPAnnualReimbursementAmt'] == 0) & (train_iobp_df['Admitted?'] == 1)]\
                   ['PotentialFraud'].value_counts()))

     PotentialFraud
Yes             249
No              164

• So, out 413 claims 60% of them means 249 are fraudulent whereas 164 are non-fraudulent.

OBSERVATION

• Thus, we can say that if the IP Annual Amount is 0 and patient was admitted to the hospital then chances of being fraudulent are high.

• Relationship b/w IPAnnualReimbursementAmt and Potentially Fraud for both the Genders

with plt.style.context('seaborn-poster'):
    fig = sns.boxplot(data=train_iobp_df[train_iobp_df['Admitted?'] == 1], x='PotentialFraud',y='IPAnnualReimbursementAmt', hue='Gender',
                      palette='flare')   
    # Providing the labels and title to the graph
    plt.xlabel("Potentially Fraud?", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Annual IP Re-Imb Amount\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Annual IP Re-Imb Amount of both the genders for Potentially Fraud & Non-Fraud Providers\n", fontdict=title_font_dict)
    plt.plot();

OBSERVATION

• The above plot clearly shows us that there is no difference in the distribution of Annual IP Re-Imb Amount of males and females for Potentially Fraud and Non-Fraud Providers.
  • Therefore, we can say that Annual IP Re-Imb Amount might not be useful in segregating the Fraud cases.

• Relationship b/w IPAnnualReimbursementAmt and Potentially Fraud for Is_Alive?

with plt.style.context('seaborn-poster'):
    fig = sns.boxplot(data=train_iobp_df[train_iobp_df['Admitted?'] == 1], x='PotentialFraud',y='IPAnnualReimbursementAmt', hue='Is_Alive?',
                      palette='prism')      
    # Providing the labels and title to the graph
    plt.xlabel("Potentially Fraud?", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Annual IP Re-Imb Amount\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Annual IP Re-Imb Amount patient life status for Potentially Fraud & Non-Fraud Providers\n", fontdict=title_font_dict)
    plt.legend(loc="upper center", title='Is_Alive?');

OBSERVATION

• The above plot clearly shows us that there is no difference in the distribution of Annual IP Re-Imb Amount on patient life status for Potentially Fraud and Non-Fraud Providers.
  • Therefore, we can say that Annual IP Re-Imb Amount might not be useful in segregating the Fraud cases.

• Relationship b/w IPAnnualReimbursementAmt and Potentially Fraud for all Human Races

with plt.style.context('seaborn-poster'):
    fig = sns.boxplot(data=train_iobp_df[train_iobp_df['Admitted?'] == 1], x='PotentialFraud',y='IPAnnualReimbursementAmt', hue='Race',
                      palette='plasma')   
    # Providing the labels and title to the graph
    plt.xlabel("Potentially Fraud?", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Annual IP Re-Imb Amount\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Annual IP Re-Imb Amount of all human races for Potentially Fraud & Non-Fraud Providers\n", fontdict=title_font_dict)
    plt.legend(loc="upper center", title="Race");

OBSERVATION

• The above plot clearly shows us that there is no difference in the distribution of Annual IP Re-Imb Amount of all human races for Potentially Fraud and Non-Fraud Providers.
  • Therefore, we can say that Annual IP Re-Imb Amount might not be useful in segregating the Fraud cases.

• Relationship b/w IPAnnualReimbursementAmt and Potentially Fraud for RenalDiseaseIndicator

with plt.style.context('seaborn-poster'):
    fig = sns.boxplot(data=train_iobp_df[train_iobp_df['Admitted?'] == 1], x='PotentialFraud',y='IPAnnualReimbursementAmt', 
                      hue='RenalDiseaseIndicator',palette='magma')   
    # Providing the labels and title to the graph
    plt.xlabel("Potentially Fraud?", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Annual IP Re-Imb Amount\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Annual IP Re-Imb Amount of whether having RKD influence Potentially Fraud & Non-Fraud Providers?\n", fontdict=title_font_dict)
    plt.legend(loc='upper center', title='RKD');

OBSERVATION

• The above plot clearly shows us that there is a very slight difference in the distribution of Annual IP Re-Imb Amount of patients with or w/o RKD for Potentially Fraud and Non-Fraud Providers.
  • Therefore, we can say that Annual IP Re-Imb Amount might not be useful in segregating the Fraud cases.

Does OPAnnualReimbursementAmt influences Potentially Fraud?

• Relationship b/w OPAnnualReimbursementAmt and Potentially Fraud

with plt.style.context('seaborn-poster'):
    fig = sns.boxenplot(data=train_iobp_df[train_iobp_df['Admitted?'] == 0], y='PotentialFraud', x='OPAnnualReimbursementAmt', 
                         palette='dark', orient='h')   
    # Providing the labels and title to the graph
    plt.ylabel("Potentially Fraud?\n", fontdict=label_font_dict)
    plt.yticks(rotation=90, fontsize=12)
    plt.xlabel("\nAnnual OP Re-Imb Amount", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.title("Annual OP Re-Imb Amount for Potentially Fraud & Non-Fraud Providers\n", fontdict=title_font_dict)
    plt.plot();

OBSERVATION

• The above plot clearly shows us that there is no difference in the distribution of Annual OP Re-Imb Amount for Potentially Fraud and Non-Fraud Providers.
  • And, it looks like a Pareto Distribution.

Why do we have OP Annual Re-Imb Amount as 0 for Admitted Patients?

print(pd.DataFrame(train_iobp_df[(train_iobp_df['OPAnnualReimbursementAmt'] == 0)]['Admitted?'].value_counts()))

   Admitted?
1       3909
0       1009

• So, we have 1009 claims where Patients were not admitted to the hospital but the allocated OP Annual Re-Imb Amt is 0.

print(pd.DataFrame(train_iobp_df[(train_iobp_df['OPAnnualReimbursementAmt'] == 0) & (train_iobp_df['Admitted?'] == 0)]\
                   ['PotentialFraud'].value_counts()))

     PotentialFraud
No              617
Yes             392

• So, out 1009 claims 38% of them means 392 are fraudulent whereas 62% or 617 are non-fraudulent.

OBSERVATION

• Thus, we can say that if the OP Annual Amount is 0 and patient was not admitted to the hospital then chances of being fraudulent are less.

• Relationship b/w OPAnnualReimbursementAmt and Potentially Fraud for both the Genders

with plt.style.context('seaborn-poster'):
    fig = sns.boxplot(data=train_iobp_df[train_iobp_df['Admitted?'] == 0], x='PotentialFraud',y='OPAnnualReimbursementAmt', hue='Gender',
                      palette='flare')   
    # Providing the labels and title to the graph
    plt.xlabel("Potentially Fraud?", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Annual OP Re-Imb Amount\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Annual OP Re-Imb Amount of both the genders for Potentially Fraud & Non-Fraud Providers\n", fontdict=title_font_dict)
    plt.plot();

OBSERVATION

• The above plot clearly shows us that there is no difference in the distribution of Annual OP Re-Imb Amount of males and females for Potentially Fraud and Non-Fraud Providers.
  • Therefore, we can say that Annual OP Re-Imb Amount might not be useful in segregating the Fraud cases.

• Relationship b/w OPAnnualReimbursementAmt and Potentially Fraud for Is_Alive?

with plt.style.context('seaborn-poster'):
    fig = sns.boxplot(data=train_iobp_df[train_iobp_df['Admitted?'] == 0], x='PotentialFraud',y='OPAnnualReimbursementAmt', hue='Is_Alive?',
                      palette='prism')      
    # Providing the labels and title to the graph
    plt.xlabel("Potentially Fraud?", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Annual OP Re-Imb Amount\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Annual OP Re-Imb Amount patient life status for Potentially Fraud & Non-Fraud Providers\n", fontdict=title_font_dict)
    plt.legend(loc="upper center", title='Is_Alive?');

OBSERVATION

• The above plot clearly shows us that there is no difference in the distribution of Annual OP Re-Imb Amount on patient life status for Potentially Fraud and Non-Fraud Providers.
  • Therefore, we can say that Annual OP Re-Imb Amount might not be useful in segregating the Fraud cases.

• Relationship b/w OPAnnualReimbursementAmt and Potentially Fraud for all Human Races

with plt.style.context('seaborn-poster'):
    fig = sns.boxplot(data=train_iobp_df[train_iobp_df['Admitted?'] == 0], x='PotentialFraud',y='OPAnnualReimbursementAmt', hue='Race',
                      palette='plasma')   
    # Providing the labels and title to the graph
    plt.xlabel("Potentially Fraud?", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Annual OP Re-Imb Amount\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Annual OP Re-Imb Amount of all human races for Potentially Fraud & Non-Fraud Providers\n", fontdict=title_font_dict)
    plt.legend(loc="upper center", title="Race");

OBSERVATION

• The above plot clearly shows us that there is no difference in the distribution of Annual OP Re-Imb Amount of all human races for Potentially Fraud and Non-Fraud Providers.
  • Therefore, we can say that Annual OP Re-Imb Amount might not be useful in segregating the Fraud cases.

• Relationship b/w OPAnnualReimbursementAmt and Potentially Fraud for RenalDiseaseIndicator

with plt.style.context('seaborn-poster'):
    fig = sns.boxplot(data=train_iobp_df[train_iobp_df['Admitted?'] == 0], x='PotentialFraud',y='OPAnnualReimbursementAmt', 
                      hue='RenalDiseaseIndicator',palette='magma')   
    # Providing the labels and title to the graph
    plt.xlabel("Potentially Fraud?", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Annual OP Re-Imb Amount\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Annual OP Re-Imb Amount of whether having RKD influence Potentially Fraud & Non-Fraud Providers?\n", fontdict=title_font_dict)
    plt.legend(loc='upper center', title='RKD');

OBSERVATION

• The above plot clearly shows us that there is a very slight difference in the distribution of Annual OP Re-Imb Amount of patients with or w/o RKD for Potentially Fraud and Non-Fraud Providers.
  • Therefore, we can say that Annual OP Re-Imb Amount might not be useful in segregating the Fraud cases.

Adding New Feature :: Total Number of false claims filed by a Provider

- Logic :: COUNT(all claims submitted by a Provider) - COUNT(all non-fraud claims submitted by a Provider)

REASONING

• The idea behind adding this feature is to introduce a way by which we can see how many fraud or non fraud claims been submitted by a provider.

• Generally what has been observed in medicare frauds is that many small hospitals from rural places had been intentionally used for filing the false claims by giving them bribes or in desire of kickbacks. Thus, for such providers total claims submitted will be less but majority of them will be false.

• But, the problem in the given dataset after joining(IP, OP, BENE with PRV TGT) is that

• if a provider was FRAUD then all the claims associated to it will be marked as FRAUD which I believe is wrong and doesn't provide any useful information.

Adding New Feature :: Total Number of claims or cases seen by Attending Physician

# Total unique number of Attended Physicians
print("Unique number of Attending Physicians present in the dataset are --> {}".format(train_iobp_df['AttendingPhysician'].nunique()))

Unique number of Attending Physicians present in the dataset are --> 82063

train_iobp_df['Att_Phy_tot_claims'] = train_iobp_df.groupby(['AttendingPhysician'])['ClaimID'].transform('count')
train_iobp_df['Att_Phy_tot_claims'].describe()

count    556703.000000
mean        138.634829
std         293.669039
min           1.000000
25%           7.000000
50%          33.000000
75%         116.000000
max        2534.000000
Name: Att_Phy_tot_claims, dtype: float64

with plt.style.context('seaborn-poster'):
    sns.kdeplot(x=train_iobp_df[train_iobp_df['PotentialFraud'] == 'Yes']['Att_Phy_tot_claims'],color='red')
    sns.kdeplot(x=train_iobp_df[train_iobp_df['PotentialFraud'] == 'No']['Att_Phy_tot_claims'],color='green')
    # Providing the labels and title to the graph
    plt.xlabel("\nAttending Physicians Total Claims Submitted", fontdict=label_font_dict)
    plt.xticks(np.arange(0,2800,100), rotation=90, fontsize=11)
    plt.minorticks_on()
    plt.title("Distribution of Total claims filed by Attending Physicians", fontdict=title_font_dict)
    plt.legend(labels=["Yes", "No"], title="Potential Fraud?");

with plt.style.context('seaborn-poster'):
    fig = sns.boxplot(data=train_iobp_df, y='PotentialFraud', x='Att_Phy_tot_claims', palette='prism_r', orient='h')   
    # Providing the labels and title to the graph
    plt.ylabel("Potentially Fraud?\n", fontdict=label_font_dict)
    plt.xticks(np.arange(0,2800,100), rotation=90, fontsize=12)
    plt.xlabel("\nAttending Physician Total Claims", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.title("Total claims filed by Attending Physicians", fontdict=title_font_dict)
    plt.plot();

OBSERVATION

• The above KDE and Box plots suggests that the newly added feature Att_Phy_tot_claims may be useful in segregating the potentially fraud and non-fraudulent cases.
  • For example, we can say that if total claims filed by a Attending Physician is greater than 500 then chances of being fraudulent are high.

Adding New Feature :: Total Number of claims or cases seen by Opearting Physician

# Total unique number of Operating Physicians
print("Unique number of Operating Physicians present in the dataset are --> {}".format(train_iobp_df['OperatingPhysician'].nunique()))

Unique number of Operating Physicians present in the dataset are --> 35315

train_iobp_df['Opr_Phy_tot_claims'] = train_iobp_df.groupby(['OperatingPhysician'])['ClaimID'].transform('count')
train_iobp_df['Opr_Phy_tot_claims'].describe()

count    114447.000000
mean         27.204811
std          52.687759
min           1.000000
25%           2.000000
50%           8.000000
75%          25.000000
max         424.000000
Name: Opr_Phy_tot_claims, dtype: float64

with plt.style.context('seaborn-poster'):
    sns.kdeplot(x=train_iobp_df[train_iobp_df['PotentialFraud'] == 'Yes']['Opr_Phy_tot_claims'],color='red')
    sns.kdeplot(x=train_iobp_df[train_iobp_df['PotentialFraud'] == 'No']['Opr_Phy_tot_claims'],color='green')
    # Providing the labels and title to the graph
    plt.xlabel("\nOperating Physicians Total Claims Submitted", fontdict=label_font_dict)
    plt.xticks(np.arange(0,500,20), rotation=90, fontsize=11)
    plt.minorticks_on()
    plt.title("Distribution of Total claims filed by Operating Physicians", fontdict=title_font_dict)
    plt.legend(labels=["Yes", "No"], title="Potential Fraud?");

with plt.style.context('seaborn-poster'):
    fig = sns.boxplot(data=train_iobp_df, y='PotentialFraud', x='Opr_Phy_tot_claims', palette='prism_r', orient='h')   
    # Providing the labels and title to the graph
    plt.ylabel("Potentially Fraud?\n", fontdict=label_font_dict)
    plt.xticks(np.arange(0,500,20), rotation=90, fontsize=12)
    plt.xlabel("\nOperating Physician Total Claims", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.title("Total claims filed by Operating Physicians", fontdict=title_font_dict)
    plt.plot();

OBSERVATION

• The above KDE and Box plots suggests that the newly added feature Opr_Phy_tot_claims may be useful in segregating the potentially fraud and non-fraudulent cases.
  • Results are very similar to Att_Phy_tot_claims.
    • For example, we can say that if total claims filed by a Operating Physician are greater than 100 then chances of being fraudulent are high.

Adding New Feature :: Total Number of claims or cases seen by Other Physician

# Total unique number of Other Physicians
print("Unique number of Other Physicians present in the dataset are --> {}".format(train_iobp_df['OtherPhysician'].nunique()))

Unique number of Other Physicians present in the dataset are --> 46457

train_iobp_df['Oth_Phy_tot_claims'] = train_iobp_df.groupby(['OtherPhysician'])['ClaimID'].transform('count')
train_iobp_df['Oth_Phy_tot_claims'].describe()

count    199736.000000
mean         90.207914
std         208.017235
min           1.000000
25%           3.000000
50%          15.000000
75%          60.000000
max        1247.000000
Name: Oth_Phy_tot_claims, dtype: float64

with plt.style.context('seaborn-poster'):
    sns.kdeplot(x=train_iobp_df[train_iobp_df['PotentialFraud'] == 'Yes']['Oth_Phy_tot_claims'],color='red')
    sns.kdeplot(x=train_iobp_df[train_iobp_df['PotentialFraud'] == 'No']['Oth_Phy_tot_claims'],color='green')
    # Providing the labels and title to the graph
    plt.xlabel("Other Physicians Total Claims Submitted", fontdict=label_font_dict)
    plt.xticks(np.arange(0,1450,50), rotation=90, fontsize=11)
    plt.minorticks_on()
    plt.title("Distribution of Total claims filed by Other Physicians", fontdict=title_font_dict)
    plt.legend(labels=["Yes", "No"], title="Potential Fraud?");

with plt.style.context('seaborn-poster'):
    fig = sns.boxplot(data=train_iobp_df, y='PotentialFraud', x='Oth_Phy_tot_claims', palette='prism_r', orient='h')   
    # Providing the labels and title to the graph
    plt.ylabel("Potentially Fraud?\n", fontdict=label_font_dict)
    plt.xticks(np.arange(0,1500,50), rotation=90, fontsize=12)
    plt.xlabel("Other Physician Total Claims", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.title("Total claims filed by Other Physicians", fontdict=title_font_dict)
    plt.plot();

OBSERVATION

• The above KDE and Box plots suggests that the newly added feature Oth_Phy_tot_claims may be useful in segregating the potentially fraud and non-fraudulent cases.
  • Results are very similar to Att_Phy_tot_claims.
    • For example, we can say that if total claims filed by a Other Physician are greater than 100 then chances of being fraudulent are high.

# Simultaneously viewing the plots for better understanding
with plt.style.context('seaborn'):
    fig, (ax1, ax2) = plt.subplots(nrows=1, ncols=2, figsize=(15,8), sharey=False)
    sns.boxplot(data=train_iobp_df[train_iobp_df["PotentialFraud"] == 'Yes'][['Att_Phy_tot_claims','Opr_Phy_tot_claims','Oth_Phy_tot_claims']],
                ax=ax1, palette='viridis')
    ax1.set_title("Potential Fraud = Yes", fontdict=title_font_dict)
    ax1.set_xlabel("\nPhysicians Categories", fontdict=label_font_dict)
    ax1.set_ylabel("Claims Filed", fontdict=label_font_dict)
    
    sns.boxplot(data=train_iobp_df[train_iobp_df["PotentialFraud"] == 'No'][['Att_Phy_tot_claims','Opr_Phy_tot_claims','Oth_Phy_tot_claims']],
                ax=ax2, palette='magma')
    ax2.set_title("Potential Fraud = No", fontdict=title_font_dict)
    ax2.set_xlabel("\nPhysicians Categories", fontdict=label_font_dict)
    ax2.set_ylabel("Claims Filed", fontdict=label_font_dict)
    # Providing the title to the figure
    fig.suptitle("Distribution of Total Claims filed by Attending(Att), Operating(Opr) and Other(Oth) Physicians.\n", fontdict=title_font_dict)
    plt.minorticks_on()
    plt.plot();    

OBSERVATION

• The above Box plots suggests that theses newly added features may be slightly useful in segregating the potentially fraud and non-fraudulent cases.
  • As we can see that the total false claims filed by physicians are slightly more than the non-false claims filed by them.

Adding Combined Feature :: Att_Opr_Oth_Phy_Tot_Claims

• It represents the total claims submitted by Attending, Operating and Other Physicians.

  • Reasoning :: The idea behind adding this feature is to see whether a total of physicians claims submission will help in influencing the potential frauds.

• Logic :: Att_Phy_tot_claims + Opr_Phy_tot_claims + Oth_Phy_tot_claims

train_iobp_df['Att_Phy_tot_claims'].fillna(value=0, inplace=True)
train_iobp_df['Opr_Phy_tot_claims'].fillna(value=0, inplace=True)
train_iobp_df['Oth_Phy_tot_claims'].fillna(value=0, inplace=True)

train_iobp_df['Att_Opr_Oth_Phy_Tot_Claims'] = train_iobp_df['Att_Phy_tot_claims'] + train_iobp_df['Opr_Phy_tot_claims'] + train_iobp_df['Oth_Phy_tot_claims']

train_iobp_df['Att_Opr_Oth_Phy_Tot_Claims'].describe()

count    558211.000000
mean        176.115666
std         379.833208
min           0.000000
25%           9.000000
50%          41.000000
75%         144.000000
max        3372.000000
Name: Att_Opr_Oth_Phy_Tot_Claims, dtype: float64

with plt.style.context('seaborn-poster'):
    sns.kdeplot(x=train_iobp_df[train_iobp_df['PotentialFraud'] == 'Yes']['Att_Opr_Oth_Phy_Tot_Claims'],color='red')
    sns.kdeplot(x=train_iobp_df[train_iobp_df['PotentialFraud'] == 'No']['Att_Opr_Oth_Phy_Tot_Claims'],color='green')
    # Providing the labels and title to the graph
    plt.xlabel("\nAttending, Operating & Other Physicians Total Claims Submitted", fontdict=label_font_dict)
    plt.xticks(np.arange(0,3600,100), rotation=90, fontsize=11)
    plt.minorticks_on()
    plt.title("Distribution of Total Claims filed by Attending, Operating & Other Physicians", fontdict=title_font_dict)
    plt.legend(labels=["Yes", "No"], title="Potential Fraud?");

with plt.style.context('seaborn-poster'):
    fig = sns.boxplot(data=train_iobp_df, y='PotentialFraud', x='Att_Opr_Oth_Phy_Tot_Claims', palette='prism_r', orient='h')   
    # Providing the labels and title to the graph
    plt.ylabel("Potentially Fraud?\n", fontdict=label_font_dict)
    plt.xticks(np.arange(0,3600,100), rotation=90, fontsize=12)
    plt.xlabel("\nAttending, Operating & Other Physicians Total Claims", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.title("Total claims filed by Attending, Operating & Other Physicians", fontdict=title_font_dict)
    plt.plot();

OBSERVATION

• The above Box plots shows us the similar results like the previous features. And, there is a slight difference in the data distributions which may be useful in segregating the potential frauds.

Adding 3 New Features :: Prv_Tot_Att_Phy, Prv_Tot_Opr_Phy and Prv_Tot_Oth_Phy

• These features will represent the total Attending, Operating and Other Physicians for every provider.
  • Reasoning :: The idea behind adding this feature is to see if a provider has wroked with very less or very high number of physicians then does that increases or decreases the chances of potential fraud.

train_iobp_df["Prv_Tot_Att_Phy"] = train_iobp_df.groupby(['Provider'])['AttendingPhysician'].transform('count')
train_iobp_df["Prv_Tot_Opr_Phy"] = train_iobp_df.groupby(['Provider'])['OperatingPhysician'].transform('count')
train_iobp_df["Prv_Tot_Oth_Phy"] = train_iobp_df.groupby(['Provider'])['OtherPhysician'].transform('count')

# Nulls in the above features
train_iobp_df.isna().sum().tail(3)

Prv_Tot_Att_Phy    0
Prv_Tot_Opr_Phy    0
Prv_Tot_Oth_Phy    0
dtype: int64

train_iobp_df["Prv_Tot_Att_Phy"].describe()

count    558211.000000
mean        820.206469
std        1271.272090
min           1.000000
25%         122.000000
50%         359.000000
75%        1013.000000
max        8207.000000
Name: Prv_Tot_Att_Phy, dtype: float64

• The average number of attending physicians for providers are 820.

train_iobp_df["Prv_Tot_Opr_Phy"].describe()

count    558211.000000
mean        155.030023
std         228.266693
min           0.000000
25%          25.000000
50%          73.000000
75%         185.000000
max        1441.000000
Name: Prv_Tot_Opr_Phy, dtype: float64

• The average number of operating physicians for providers are 155.

train_iobp_df["Prv_Tot_Oth_Phy"].describe()

count    558211.000000
mean        306.781194
std         497.246984
min           0.000000
25%          37.000000
50%         120.000000
75%         381.000000
max        3209.000000
Name: Prv_Tot_Oth_Phy, dtype: float64

• The average number of other physicians for providers are 306.

• Relationship b/w PROVIDER and ATTENDING PHYSICIAN

with plt.style.context('seaborn-poster'):
    sns.kdeplot(x=train_iobp_df[train_iobp_df['PotentialFraud'] == 'Yes']['Prv_Tot_Att_Phy'],color='red')
    sns.kdeplot(x=train_iobp_df[train_iobp_df['PotentialFraud'] == 'No']['Prv_Tot_Att_Phy'],color='blue')
    # Providing the labels and title to the graph
    plt.xlabel("\nProviders interacted with how many attending physicians?", fontdict=label_font_dict)
    plt.xticks(np.arange(0,8800,400), rotation=90, fontsize=11)
    plt.minorticks_on()
    plt.title("Distribution of Providers interaction with attending physicians", fontdict=title_font_dict)
    plt.legend(labels=["Yes", "No"], title="Potential Fraud?");

with plt.style.context('seaborn-poster'):
    fig = sns.boxplot(data=train_iobp_df, y='PotentialFraud', x='Prv_Tot_Att_Phy', palette='prism_r', orient='h')   
    # Providing the labels and title to the graph
    plt.ylabel("Potentially Fraud?\n", fontdict=label_font_dict)
    plt.xticks(np.arange(0,8800,400), rotation=90, fontsize=12)
    plt.xlabel("\nProviders interacted with how many attending physicians?", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.title("Distribution of Providers interaction with attending physicians", fontdict=title_font_dict)
    plt.plot();

OBSERVATION

• The above KDE and Box plots are quite interesting as we can see that if Prv_Tot_Att_Phy is high then chances of fraud is quite high.

• Relationship b/w PROVIDER and OPERATING PHYSICIAN

with plt.style.context('seaborn-poster'):
    sns.kdeplot(x=train_iobp_df[train_iobp_df['PotentialFraud'] == 'Yes']['Prv_Tot_Opr_Phy'],color='red')
    sns.kdeplot(x=train_iobp_df[train_iobp_df['PotentialFraud'] == 'No']['Prv_Tot_Opr_Phy'],color='blue')
    # Providing the labels and title to the graph
    plt.xlabel("\nProviders interacted with how many operating physicians?", fontdict=label_font_dict)
    plt.xticks(np.arange(0,1600,100), rotation=90, fontsize=11)
    plt.minorticks_on()
    plt.title("Distribution of Providers interaction with operating physicians", fontdict=title_font_dict)
    plt.legend(labels=["Yes", "No"], title="Potential Fraud?");

with plt.style.context('seaborn-poster'):
    fig = sns.boxplot(data=train_iobp_df, y='PotentialFraud', x='Prv_Tot_Opr_Phy', palette='prism_r', orient='h')   
    # Providing the labels and title to the graph
    plt.ylabel("Potentially Fraud?\n", fontdict=label_font_dict)
    plt.xticks(np.arange(0,1600,100), rotation=90, fontsize=12)
    plt.xlabel("\nProviders interacted with how many operating physicians?", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.title("Distribution of Providers interaction with operating physicians", fontdict=title_font_dict)
    plt.plot();

OBSERVATION

• The above KDE and Box plots are quite interesting as we can see that if Prv_Tot_Opr_Phy is high then chances of fraud is quite high.

• Relationship b/w PROVIDER and OTHER PHYSICIAN

with plt.style.context('seaborn-poster'):
    sns.kdeplot(x=train_iobp_df[train_iobp_df['PotentialFraud'] == 'Yes']['Prv_Tot_Oth_Phy'],color='red')
    sns.kdeplot(x=train_iobp_df[train_iobp_df['PotentialFraud'] == 'No']['Prv_Tot_Oth_Phy'],color='blue')
    # Providing the labels and title to the graph
    plt.xlabel("\nProviders interacted with how many other physicians?", fontdict=label_font_dict)
    plt.xticks(np.arange(0,3600,200), rotation=90, fontsize=11)
    plt.minorticks_on()
    plt.title("Distribution of Providers interaction with other physicians", fontdict=title_font_dict)
    plt.legend(labels=["Yes", "No"], title="Potential Fraud?");

with plt.style.context('seaborn-poster'):
    fig = sns.boxplot(data=train_iobp_df, y='PotentialFraud', x='Prv_Tot_Oth_Phy', palette='prism_r', orient='h')   
    # Providing the labels and title to the graph
    plt.ylabel("Potentially Fraud?\n", fontdict=label_font_dict)
    plt.xticks(np.arange(0,3600,200), rotation=90, fontsize=12)
    plt.xlabel("\nProviders interacted with how many other physicians?", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.title("Distribution of Providers interaction with other physicians", fontdict=title_font_dict)
    plt.plot();

OBSERVATION

• The above KDE and Box plots are quite interesting as we can see that if Prv_Tot_Oth_Phy is high then chances of fraud is quite high.

Adding Combined Feature :: Prv_Tot_Att_Opr_Oth_Phys

• It represents the total of all kind of physicians that a provider has interacted with.

  • Reasoning :: The idea behind adding this feature is to see whether a fraudulent provider interacts with higher or lower numberof of various physicians.

• Logic :: Prv_Tot_Att_Phy + Prv_Tot_Opr_Phy + Prv_Tot_Oth_Phy

train_iobp_df['Prv_Tot_Att_Opr_Oth_Phys'] = train_iobp_df['Prv_Tot_Att_Phy'] + train_iobp_df['Prv_Tot_Opr_Phy'] + train_iobp_df['Prv_Tot_Oth_Phy']

with plt.style.context('seaborn-poster'):
    sns.kdeplot(x=train_iobp_df[train_iobp_df['PotentialFraud'] == 'Yes']['Prv_Tot_Att_Opr_Oth_Phys'],color='red')
    sns.kdeplot(x=train_iobp_df[train_iobp_df['PotentialFraud'] == 'No']['Prv_Tot_Att_Opr_Oth_Phys'],color='blue')
    # Providing the labels and title to the graph
    plt.xlabel("\nProviders interacted with how many all kind of physicians?", fontdict=label_font_dict)
    plt.xticks(np.arange(0,14000,1000), rotation=90, fontsize=11)
    plt.minorticks_on()
    plt.title("Distribution of Providers interaction with all kind of physicians", fontdict=title_font_dict)
    plt.legend(labels=["Yes", "No"], title="Potential Fraud?");

with plt.style.context('seaborn-poster'):
    fig = sns.boxplot(data=train_iobp_df, y='PotentialFraud', x='Prv_Tot_Att_Opr_Oth_Phys', palette='prism_r', orient='h')   
    # Providing the labels and title to the graph
    plt.ylabel("Potentially Fraud?\n", fontdict=label_font_dict)
    plt.xticks(np.arange(0,14000,1000), rotation=90, fontsize=12)
    plt.xlabel("\nProviders interacted with how many all kind of physicians?", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.title("Distribution of Providers interaction with all kind of physicians", fontdict=title_font_dict)
    plt.plot();

OBSERVATION

• The above KDE and Box plots are quite interesting as we can see that if Prv_Tot_Att_Opr_Oth_Phys is high then chances of fraud is quite high.

Adding New Feature :: Total Unique Claim Admit Codes used by a PROVIDER

• Reasoning :: The idea behind adding this feature is to see how many unique number of Claim Admit Diagnosis Codes used by the Provider.
  • As there may be a pattern that if a provider has used so many Admit Diagnosis Codes then it might increases or decreases the chances of fraud.

train_iobp_df['PRV_Tot_Admit_DCodes'] = train_iobp_df.groupby(['Provider'])['ClmAdmitDiagnosisCode'].transform('nunique')

with plt.style.context('seaborn-poster'):
    sns.kdeplot(x=train_iobp_df[train_iobp_df['PotentialFraud'] == 'Yes']['PRV_Tot_Admit_DCodes'], color='red')
    sns.kdeplot(x=train_iobp_df[train_iobp_df['PotentialFraud'] == 'No']['PRV_Tot_Admit_DCodes'], color='blue')
    # Providing the labels and title to the graph
    plt.xlabel("\nProviders used unique Claim Admit Diagnosis Codes", fontdict=label_font_dict)
    plt.xticks(np.arange(0,600,50), rotation=90, fontsize=11)
    plt.minorticks_on()
    plt.title("Distribution of Providers used unique Claim Admit Diagnosis Codes", fontdict=title_font_dict)
    plt.legend(labels=["Yes", "No"], title="Potential Fraud?");

with plt.style.context('seaborn-poster'):
    fig = sns.boxplot(data=train_iobp_df, y='PotentialFraud', x='PRV_Tot_Admit_DCodes', palette='prism_r', orient='h')   
    # Providing the labels and title to the graph
    plt.ylabel("Potentially Fraud?\n", fontdict=label_font_dict)
    plt.xticks(np.arange(0,600,50), rotation=90, fontsize=12)
    plt.xlabel("\nProviders used unique Claim Admit Diagnosis Codes", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.title("Distribution of Providers used unique Claim Admit Diagnosis Codes", fontdict=title_font_dict)
    plt.plot();

OBSERVATION

• The above KDE and Box plots are very interesting as we can see that if PRV_Tot_Admit_DCodes is high then chances of fraud also increases.

NOTE :: What didn't worked?

• I also looked to add the unique number of Admit Diagnosis Codes used by the 3 different class of physicians but the variation was very minimal, thus not added those features.

Adding New Feature :: Total Unique Number of Diagnosis Group Codes used by a PROVIDER

• Reasoning :: The idea behind adding this feature is to see how many unique Diagnosis Group Codes used by the Provider.
  • As there may be a pattern that if a provider has used so many Diagnosis Group Codes then it might increases or decreases the chances of fraud.

train_iobp_df['PRV_Tot_DGrpCodes'] = train_iobp_df.groupby(['Provider'])['DiagnosisGroupCode'].transform('nunique')

with plt.style.context('seaborn-poster'):
    sns.kdeplot(x=train_iobp_df[train_iobp_df['PotentialFraud'] == 'Yes']['PRV_Tot_DGrpCodes'], color='red')
    sns.kdeplot(x=train_iobp_df[train_iobp_df['PotentialFraud'] == 'No']['PRV_Tot_DGrpCodes'], color='blue')
    # Providing the labels and title to the graph
    plt.xlabel("\nProviders used unique Diagnosis Group Codes", fontdict=label_font_dict)
    plt.xticks(np.arange(0,400,40), rotation=90, fontsize=11)
    plt.minorticks_on()
    plt.title("Distribution of Providers used unique Diagnosis Group Codes", fontdict=title_font_dict)
    plt.legend(labels=["Yes", "No"], title="Potential Fraud?");

with plt.style.context('seaborn-poster'):
    fig = sns.boxplot(data=train_iobp_df, y='PotentialFraud', x='PRV_Tot_DGrpCodes', palette='prism_r', orient='h')   
    # Providing the labels and title to the graph
    plt.ylabel("Potentially Fraud?\n", fontdict=label_font_dict)
    plt.xticks(np.arange(0,400,40), rotation=90, fontsize=12)
    plt.xlabel("\nProviders used unique Diagnosis Group Codes", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.title("Distribution of Providers used unique Diagnosis Group Codes", fontdict=title_font_dict)
    plt.plot();

OBSERVATION

• The above KDE and Box plots suggest that if PRV_Tot_Admit_DCodes is high then it slightly increases the chances of fraud.

NOTE :: What didn't worked?

• I also looked to add the unique number of Diagnosis Group Codes used by the 3 different class of physicians but the variation was very minimal, thus not added those features.

NOTE :: What didn't worked?

• I also looked to add in how many claims a unique Diagnosis Group Code is used but there was no variation at all, thus not added that feature. Kindly refer to the below image:

NOTE :: What didn't worked?

• I also looked to add DOB -- Month, DOD -- Year and DOD -- Month in order to see whether we can find some pattern of bogus DOB or DOD but there was no variation at all, thus not added that feature. Also, raw DOB -- Year also showed no variation. Kindly refer to the below image:

Adding New Feature :: Total unique Date of Birth years of beneficiaries provided by a Provider

• Reasoning :: The idea behind adding this feature is that if a provider has very high variability in the year of birth of patients then that might be one of the signs of medicare frauds.
  • Because generally private hospitals who treat poor patients make false claims on their names. For example, Nazia is 10 years old. But, according to a claim filed by Chhattisgarh-based Shaheed Hospital with the Rashtriya Swasthya Bima Yojna (RSBY), she has delivered a baby after a caesarean operation. Mukul (name changed) is only 7. But Agarwal Hospital, Raipur, has made a claim for removing cataract from his eyes.

Read more at: https://economictimes.indiatimes.com/news/politics-and-nation/private-hospitals-perform-fake-surgeries-to-claim-thousands-in-insurance-cover/articleshow/16934229.cms?utm_source=contentofinterest&utm_medium=text&utm_campaign=cppst

train_iobp_df['DOB_Year'] = train_iobp_df['DOB'].dt.year

train_iobp_df['PRV_Tot_Unq_DOB_Years'] = train_iobp_df.groupby(['Provider'])['DOB_Year'].transform('nunique')

train_iobp_df['PRV_Tot_Unq_DOB_Years'].describe()

count    558211.000000
mean         50.615590
std          18.190988
min           1.000000
25%          38.000000
50%          54.000000
75%          67.000000
max          75.000000
Name: PRV_Tot_Unq_DOB_Years, dtype: float64

with plt.style.context('seaborn-poster'):
    sns.kdeplot(x=train_iobp_df[train_iobp_df['PotentialFraud'] == 'Yes']['PRV_Tot_Unq_DOB_Years'], color='red')
    sns.kdeplot(x=train_iobp_df[train_iobp_df['PotentialFraud'] == 'No']['PRV_Tot_Unq_DOB_Years'], color='blue')
    # Providing the labels and title to the graph
    plt.xlabel("\nTotal unique Years of birth of patients", fontdict=label_font_dict)
    plt.xticks(np.arange(0,80,4), rotation=90, fontsize=11)
    plt.minorticks_on()
    plt.title("Distribution Providers treated Patients of various DOB Years", fontdict=title_font_dict)
    plt.legend(labels=["Yes", "No"], title="Potential Fraud?");

with plt.style.context('seaborn-poster'):
    fig = sns.boxplot(data=train_iobp_df, y='PotentialFraud', x='PRV_Tot_Unq_DOB_Years', palette='prism_r', orient='h')   
    # Providing the labels and title to the graph
    plt.ylabel("Potentially Fraud?\n", fontdict=label_font_dict)
    plt.xticks(np.arange(0,80,4), rotation=90, fontsize=12)
    plt.xlabel("\nTotal unique Years of birth of patients", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.title("Distribution Providers treated Patients of various DOB Years", fontdict=title_font_dict)
    plt.plot();

OBSERVATION

• The above KDE and Box plots suggest that if PRV_Tot_Unq_DOB_Years is very high than then it increases the chances of fraud as well.
  • As shown below, if PRV_Tot_Unq_DOB_Years is greater than or equals to 67 then the very high majority of cases are fraud.

train_iobp_df[train_iobp_df['PRV_Tot_Unq_DOB_Years'] >=67]['PotentialFraud'].value_counts()

Yes    118873
No      22203
Name: PotentialFraud, dtype: int64

Adding New Feature :: Sum of patients age treated by a Provider

• Reasoning :: The idea behind adding this feature is that there might be a pattern like if the sum of patients age treated by a provider is very high or low then it might influence the fraud.

train_iobp_df['PRV_Bene_Age_Sum'] = train_iobp_df.groupby(['Provider'])['Bene_Age'].transform('sum')

train_iobp_df['PRV_Bene_Age_Sum'].describe()

count    558211.000000
mean      60903.124044
std       95028.202759
min          34.300000
25%        9007.200000
50%       26310.800000
75%       74869.500000
max      617454.100000
Name: PRV_Bene_Age_Sum, dtype: float64

with plt.style.context('seaborn-poster'):
    sns.kdeplot(x=train_iobp_df[train_iobp_df['PotentialFraud'] == 'Yes']['PRV_Bene_Age_Sum'], color='red')
    sns.kdeplot(x=train_iobp_df[train_iobp_df['PotentialFraud'] == 'No']['PRV_Bene_Age_Sum'], color='blue')
    # Providing the labels and title to the graph
    plt.xlabel("\nSum of patients age treated by Providers", fontdict=label_font_dict)
    plt.xticks(np.arange(0,620000,50000), rotation=90, fontsize=11)
    plt.minorticks_on()
    plt.title("Distribution of Sum of patients age treated by Providers", fontdict=title_font_dict)
    plt.legend(labels=["Yes", "No"], title="Potential Fraud?");

with plt.style.context('seaborn-poster'):
    fig = sns.boxplot(data=train_iobp_df, y='PotentialFraud', x='PRV_Bene_Age_Sum', palette='prism_r', orient='h')   
    # Providing the labels and title to the graph
    plt.ylabel("Potentially Fraud?\n", fontdict=label_font_dict)
    plt.xticks(np.arange(0,620000,50000), rotation=90, fontsize=11)
    plt.xlabel("\nSum of patients age treated by Providers", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.title("Distribution of Sum of patients age treated by Providers", fontdict=title_font_dict)
    plt.plot();

OBSERVATION

• The above KDE and Box plots suggest that if PRV_Bene_Age_Sum is high then it increases the chances of fraud.

Adding New Feature :: Sum of Insc Claim Re-Imb Amount for a Provider

• Reasoning :: The idea behind adding this feature is that there might be a pattern like if the sum of claim re-imb amount for a provider is very high or low then it might influence the fraud.

train_iobp_df['PRV_Insc_Clm_ReImb_Amt'] = train_iobp_df.groupby(['Provider'])['InscClaimAmtReimbursed'].transform('sum')

train_iobp_df['PRV_Insc_Clm_ReImb_Amt'].describe()

count    5.582110e+05
mean     4.877429e+05
std      7.367223e+05
min      0.000000e+00
25%      6.369000e+04
50%      2.036000e+05
75%      5.969000e+05
max      5.996050e+06
Name: PRV_Insc_Clm_ReImb_Amt, dtype: float64

with plt.style.context('seaborn-poster'):
    sns.kdeplot(x=train_iobp_df[train_iobp_df['PotentialFraud'] == 'Yes']['PRV_Insc_Clm_ReImb_Amt'], color='red')
    sns.kdeplot(x=train_iobp_df[train_iobp_df['PotentialFraud'] == 'No']['PRV_Insc_Clm_ReImb_Amt'], color='blue')
    # Providing the labels and title to the graph
    plt.xlabel("\nSum of Insc Claim Re-Imb Amount for a Provider", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.title("Distribution of Sum of Insc Claim Re-Imb Amount for a Provider", fontdict=title_font_dict)
    plt.legend(labels=["Yes", "No"], title="Potential Fraud?");

with plt.style.context('seaborn-poster'):
    fig = sns.boxplot(data=train_iobp_df, y='PotentialFraud', x='PRV_Insc_Clm_ReImb_Amt', palette='prism_r', orient='h')   
    # Providing the labels and title to the graph
    plt.ylabel("Potentially Fraud?\n", fontdict=label_font_dict)
    plt.xlabel("\nSum of Insc Claim Re-Imb Amount for a Provider", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.title("Distribution of Sum of Insc Claim Re-Imb Amount for a Provider", fontdict=title_font_dict)
    plt.plot();

OBSERVATION

• The above KDE and Box plots suggest that if PRV_Insc_Clm_ReImb_Amt is high then it increases the chances of fraud.

Adding New Feature :: Total number of RKD Patients seen by a Provider

• Reasoning :: The idea behind adding this feature is that there might be a pattern like if the total number of RKD Patients seen by a Provider is very high or low then it might influence the fraud.

train_iobp_df['RenalDiseaseIndicator'] = train_iobp_df['RenalDiseaseIndicator'].apply(lambda val: 1 if val == "Y" else 0)

train_iobp_df['PRV_Tot_RKD_Patients'] = train_iobp_df.groupby(['Provider'])['RenalDiseaseIndicator'].transform('sum')

train_iobp_df['PRV_Tot_RKD_Patients'].describe()

count    558211.000000
mean        157.902616
std         233.828365
min           0.000000
25%          24.000000
50%          73.000000
75%         192.000000
max        1447.000000
Name: PRV_Tot_RKD_Patients, dtype: float64

with plt.style.context('seaborn-poster'):
    sns.kdeplot(x=train_iobp_df[train_iobp_df['PotentialFraud'] == 'Yes']['PRV_Tot_RKD_Patients'], color='red')
    sns.kdeplot(x=train_iobp_df[train_iobp_df['PotentialFraud'] == 'No']['PRV_Tot_RKD_Patients'], color='blue')
    # Providing the labels and title to the graph
    plt.xlabel("\nRKD Patients seen by a Provider", fontdict=label_font_dict)
    plt.xticks(np.arange(0,1600,100), rotation=90, fontsize=11)
    plt.minorticks_on()
    plt.title("Distribution of total number of RKD Patients seen by a Provider", fontdict=title_font_dict)
    plt.legend(labels=["Yes", "No"], title="Potential Fraud?");

with plt.style.context('seaborn-poster'):
    fig = sns.boxplot(data=train_iobp_df, y='PotentialFraud', x='PRV_Tot_RKD_Patients', palette='prism_r', orient='h')   
    # Providing the labels and title to the graph
    plt.ylabel("Potentially Fraud?\n", fontdict=label_font_dict)
    plt.xticks(np.arange(0,1600,100), rotation=90, fontsize=11)
    plt.xlabel("\nRKD Patients seen by a Provider", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.title("Distribution of total number of RKD Patients seen by a Provider", fontdict=title_font_dict)
    plt.plot();

OBSERVATION

• The above KDE and Box plots suggest that if PRV_Tot_RKD_Patients is high then it increases the chances of fraud.

Exploratory Data Analysis

Let's find some trends

Q1. Which are the Top-25 Providers with maximum number of fraudulent cases?

tmp = pd.DataFrame(train_iobp_df.groupby(['Provider','PotentialFraud'])['BeneID'].count()).reset_index()
tmp.columns = ['Provider', 'Fraud?', 'Num_of_cases']
tot_fraud_cases = tmp[tmp['Fraud?'] == 'Yes']['Num_of_cases'].sum()
tot_non_fraud_cases = tmp[tmp['Fraud?'] == 'No']['Num_of_cases'].sum()
tmp['Cases'] = tmp['Fraud?'].apply(lambda val: tot_non_fraud_cases if val == "No" else tot_fraud_cases)
tmp['Percentage'] = round(((tmp['Num_of_cases'] / tmp['Cases']) * 100),2)

tmp.head()

	Provider	Fraud?	Num_of_cases	Cases	Percentage
0	PRV51001	No	25	345415	0.01
1	PRV51003	Yes	132	212796	0.06
2	PRV51004	No	149	345415	0.04
3	PRV51005	Yes	1165	212796	0.55
4	PRV51007	No	72	345415	0.02

tmp_only_frauds = tmp[tmp['Fraud?'] == 'Yes'].sort_values(by=['Percentage'], ascending=False).reset_index(drop=True)

print(tmp_only_frauds[['Provider','Num_of_cases','Percentage']].head(25), "\n")

with plt.style.context('seaborn'):
    plt.figure(figsize=(14,8))
    fig = sns.barplot(data=tmp_only_frauds.iloc[0:25], x="Provider", y="Num_of_cases", palette='Accent')
    # Using the "patches" function we will get the location of the rectangle bars from the graph.
    ## Then by using those location(width & height) values we will add the annotations
    for p in fig.patches:
        width = p.get_width()
        height = p.get_height()
        x, y = p.get_xy()
        fig.annotate(f'{str(round((height*100)/tot_fraud_cases,2))+"%"}', (x + width/2, y + height*1.025), ha='center', fontsize=13.5, rotation=90)
    
    # Providing the labels and title to the graph
    plt.xlabel("\nTop Fraudulent Providers", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Number (or % share) of Cases\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Top-25 Providers with most number of fraudulent cases\n", fontdict=title_font_dict)
    plt.plot();

    Provider  Num_of_cases  Percentage
0   PRV51459          8240        3.87
1   PRV53797          4739        2.23
2   PRV51574          4444        2.09
3   PRV53918          3588        1.69
4   PRV54895          3436        1.61
5   PRV55215          3393        1.59
6   PRV52064          2844        1.34
7   PRV56011          2833        1.33
8   PRV55004          2399        1.13
9   PRV56560          2313        1.09
10  PRV57306          2315        1.09
11  PRV52030          2275        1.07
12  PRV52649          2156        1.01
13  PRV54772          2115        0.99
14  PRV52628          2098        0.99
15  PRV51369          2083        0.98
16  PRV51347          2067        0.97
17  PRV55039          2058        0.97
18  PRV57103          2049        0.96
19  PRV52019          1961        0.92
20  PRV51480          1924        0.90
21  PRV55462          1907        0.90
22  PRV52041          1885        0.89
23  PRV55467          1896        0.89
24  PRV54742          1892        0.89 

OBSERVATION

• The above plot shows us the Top-25 Providers with most percentage of Fraudulent Case Submissions.
  • Here, PRV51459 has the highest percentage share of fraudulent cases. The, difference b/w others providers is not that high.

Q2. Which are the Top-25 Providers with maximum number of non-fraudulent cases?

tmp_only_non_frauds = tmp[tmp['Fraud?'] == 'No'].sort_values(by=['Percentage'], ascending=False).reset_index(drop=True)

print(tmp_only_non_frauds[['Provider','Num_of_cases','Percentage']].head(25), "\n")

with plt.style.context('seaborn'):
    plt.figure(figsize=(14,8))
    fig = sns.barplot(data=tmp_only_non_frauds.iloc[0:25], x="Provider", y="Num_of_cases", palette='Accent')
    # Using the "patches" function we will get the location of the rectangle bars from the graph.
    ## Then by using those location(width & height) values we will add the annotations
    for p in fig.patches:
        width = p.get_width()
        height = p.get_height()
        x, y = p.get_xy()
        fig.annotate(f'{str(round((height*100)/tot_non_fraud_cases,2))+"%"}', (x + width/2, y + height*1.025), ha='center', fontsize=13.5, rotation=90)
    
    # Providing the labels and title to the graph
    plt.xlabel("\nTop Non-Fraudulent Providers", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Number (or % share) of Cases\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Top-25 Providers with most number of non-fraudulent cases\n", fontdict=title_font_dict)
    plt.plot();

    Provider  Num_of_cases  Percentage
0   PRV53750          1245        0.36
1   PRV55552          1206        0.35
2   PRV53394          1215        0.35
3   PRV53871          1220        0.35
4   PRV52001          1177        0.34
5   PRV52104          1189        0.34
6   PRV56559          1113        0.32
7   PRV56006          1090        0.32
8   PRV54813          1056        0.31
9   PRV52631          1057        0.31
10  PRV51509          1030        0.30
11  PRV56270          1043        0.30
12  PRV54332           995        0.29
13  PRV52605          1000        0.29
14  PRV53702          1010        0.29
15  PRV57333          1004        0.29
16  PRV55525           933        0.27
17  PRV56248           934        0.27
18  PRV57348           920        0.27
19  PRV52395           912        0.26
20  PRV56243           909        0.26
21  PRV52859           911        0.26
22  PRV53700           913        0.26
23  PRV55510           861        0.25
24  PRV54374           854        0.25 

OBSERVATION

• The above plot shows us the Top-25 Providers with most percentage of Non-Fraudulent Case Submissions.
  • Here, PRV53750 has the highest percentage share of non-fraudulent cases. But, the difference with other providers is not so high.

Q3. Which are the Top-25 Attending Physicians with maximum number of fraudulent cases?

tmp = pd.DataFrame(train_iobp_df.groupby(['AttendingPhysician','PotentialFraud'])['BeneID'].count()).reset_index()
tmp.columns = ['AttendingPhysician', 'Fraud?', 'Num_of_cases']
tot_fraud_cases = tmp[tmp['Fraud?'] == 'Yes']['Num_of_cases'].sum()
tot_non_fraud_cases = tmp[tmp['Fraud?'] == 'No']['Num_of_cases'].sum()
tmp['Cases'] = tmp['Fraud?'].apply(lambda val: tot_non_fraud_cases if val == "No" else tot_fraud_cases)
tmp['Percentage'] = round(((tmp['Num_of_cases'] / tmp['Cases']) * 100),2)

tmp.head()

	AttendingPhysician	Fraud?	Num_of_cases	Cases	Percentage
0	PHY311001	No	2	344471	0.0
1	PHY311002	Yes	1	212232	0.0
2	PHY311004	No	2	344471	0.0
3	PHY311005	No	2	344471	0.0
4	PHY311006	No	1	344471	0.0

tmp_only_frauds = tmp[tmp['Fraud?'] == 'Yes'].sort_values(by=['Percentage'], ascending=False).reset_index(drop=True)

print(tmp_only_frauds[['AttendingPhysician','Num_of_cases','Percentage']].head(25), "\n")

with plt.style.context('seaborn'):
    plt.figure(figsize=(14,8))
    fig = sns.barplot(data=tmp_only_frauds.iloc[0:25], x="AttendingPhysician", y="Num_of_cases", palette='Accent')
    # Using the "patches" function we will get the location of the rectangle bars from the graph.
    ## Then by using those location(width & height) values we will add the annotations
    for p in fig.patches:
        width = p.get_width()
        height = p.get_height()
        x, y = p.get_xy()
        fig.annotate(f'{str(round((height*100)/tot_fraud_cases,2))+"%"}', (x + width/2, y + height*1.025), ha='center', fontsize=13.5, rotation=90)
    
    # Providing the labels and title to the graph
    plt.xlabel("\nTop Fraudulent AttendingPhysician", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Number (or % share) of Cases\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Top-25 AttendingPhysician with most number of fraudulent cases\n", fontdict=title_font_dict)
    plt.plot();

   AttendingPhysician  Num_of_cases  Percentage
0           PHY330576          2534        1.19
1           PHY350277          1628        0.77
2           PHY412132          1321        0.62
3           PHY423534          1223        0.58
4           PHY314027          1200        0.57
5           PHY327046          1181        0.56
6           PHY338032          1158        0.55
7           PHY357120          1156        0.54
8           PHY337425          1156        0.54
9           PHY341578          1133        0.53
10          PHY432650          1093        0.52
11          PHY347064          1076        0.51
12          PHY344389          1000        0.47
13          PHY383481          1005        0.47
14          PHY415321          1002        0.47
15          PHY433436           924        0.44
16          PHY375453           880        0.41
17          PHY387126           762        0.36
18          PHY357307           737        0.35
19          PHY318667           711        0.34
20          PHY424712           693        0.33
21          PHY347780           678        0.32
22          PHY313278           674        0.32
23          PHY323447           664        0.31
24          PHY333735           634        0.30 

OBSERVATION

• The above plot shows us the Top-25 Attenting Physicians with most percentage of Fraudulent Case Submissions.
  • Here, PHY330576 has the highest percentage share of fraudulent cases. But, the difference b/w other physicians is not so high.

Q4. Which are the Top-25 Attenting Physicians with maximum number of non-fraudulent cases?

tmp_only_non_frauds = tmp[tmp['Fraud?'] == 'No'].sort_values(by=['Percentage'], ascending=False).reset_index(drop=True)

print(tmp_only_non_frauds[['AttendingPhysician','Num_of_cases','Percentage']].head(25), "\n")

with plt.style.context('seaborn'):
    plt.figure(figsize=(14,8))
    fig = sns.barplot(data=tmp_only_non_frauds.iloc[0:25], x="AttendingPhysician", y="Num_of_cases", palette='Accent')
    # Using the "patches" function we will get the location of the rectangle bars from the graph.
    ## Then by using those location(width & height) values we will add the annotations
    for p in fig.patches:
        width = p.get_width()
        height = p.get_height()
        x, y = p.get_xy()
        fig.annotate(f'{str(round((height*100)/tot_non_fraud_cases,2))+"%"}', (x + width/2, y + height*1.025), ha='center', fontsize=13.5, rotation=90)
    
    # Providing the labels and title to the graph
    plt.xlabel("\nTop Non-Fraudulent AttendingPhysician", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Number (or % share) of Cases\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Top-25 AttendingPhysician with most number of non-fraudulent cases\n", fontdict=title_font_dict)
    plt.plot();

   AttendingPhysician  Num_of_cases  Percentage
0           PHY351121          1053        0.31
1           PHY375943           912        0.26
2           PHY432614           716        0.21
3           PHY389456           673        0.20
4           PHY326984           686        0.20
5           PHY362889           674        0.20
6           PHY373032           618        0.18
7           PHY367255           634        0.18
8           PHY356444           600        0.17
9           PHY360179           583        0.17
10          PHY405720           544        0.16
11          PHY342223           503        0.15
12          PHY387900           514        0.15
13          PHY430054           513        0.15
14          PHY361063           470        0.14
15          PHY326049           445        0.13
16          PHY388040           447        0.13
17          PHY351973           426        0.12
18          PHY318242           422        0.12
19          PHY328307           402        0.12
20          PHY403755           429        0.12
21          PHY322775           372        0.11
22          PHY374226           363        0.11
23          PHY424939           347        0.10
24          PHY366914           333        0.10 

OBSERVATION

• The above plot shows us the Top-25 Attenting Physicians with most percentage of Non-Fraudulent Case Submissions.
  • Here, PHY351121 has the highest percentage share of non-fraudulent cases. But, the difference with other providers is not so high.

Q5. Which are the Top-25 Operating Physicians with maximum number of fraudulent cases?

tmp = pd.DataFrame(train_iobp_df.groupby(['OperatingPhysician','PotentialFraud'])['BeneID'].count()).reset_index()
tmp.columns = ['OperatingPhysician', 'Fraud?', 'Num_of_cases']
tot_fraud_cases = tmp[tmp['Fraud?'] == 'Yes']['Num_of_cases'].sum()
tot_non_fraud_cases = tmp[tmp['Fraud?'] == 'No']['Num_of_cases'].sum()
tmp['Cases'] = tmp['Fraud?'].apply(lambda val: tot_non_fraud_cases if val == "No" else tot_fraud_cases)
tmp['Percentage'] = round(((tmp['Num_of_cases'] / tmp['Cases']) * 100),2)

tmp.head()

	OperatingPhysician	Fraud?	Num_of_cases	Cases	Percentage
0	PHY311005	No	1	67497	0.00
1	PHY311010	No	1	67497	0.00
2	PHY311011	Yes	5	46950	0.01
3	PHY311014	No	3	67497	0.00
4	PHY311018	No	2	67497	0.00

tmp_only_frauds = tmp[tmp['Fraud?'] == 'Yes'].sort_values(by=['Percentage'], ascending=False).reset_index(drop=True)

print(tmp_only_frauds[['OperatingPhysician','Num_of_cases','Percentage']].head(25), "\n")

with plt.style.context('seaborn'):
    plt.figure(figsize=(14,8))
    fig = sns.barplot(data=tmp_only_frauds.iloc[0:25], x="OperatingPhysician", y="Num_of_cases", palette='Accent')
    # Using the "patches" function we will get the location of the rectangle bars from the graph.
    ## Then by using those location(width & height) values we will add the annotations
    for p in fig.patches:
        width = p.get_width()
        height = p.get_height()
        x, y = p.get_xy()
        fig.annotate(f'{str(round((height*100)/tot_fraud_cases,2))+"%"}', (x + width/2, y + height*1.025), ha='center', fontsize=13.5, rotation=90)
    
    # Providing the labels and title to the graph
    plt.xlabel("\nTop Fraudulent OperatingPhysician", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Number (or % share) of Cases\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Top-25 OperatingPhysician with most number of fraudulent cases\n", fontdict=title_font_dict)
    plt.plot();

   OperatingPhysician  Num_of_cases  Percentage
0           PHY330576           424        0.90
1           PHY424897           293        0.62
2           PHY314027           256        0.55
3           PHY423534           250        0.53
4           PHY357120           249        0.53
5           PHY412132           245        0.52
6           PHY327046           236        0.50
7           PHY381249           231        0.49
8           PHY333735           232        0.49
9           PHY341578           224        0.48
10          PHY429430           225        0.48
11          PHY337425           226        0.48
12          PHY383481           191        0.41
13          PHY347064           189        0.40
14          PHY432650           178        0.38
15          PHY344389           171        0.36
16          PHY415321           165        0.35
17          PHY433436           159        0.34
18          PHY341560           154        0.33
19          PHY387026           143        0.30
20          PHY387126           138        0.29
21          PHY357307           127        0.27
22          PHY404832           127        0.27
23          PHY347780           121        0.26
24          PHY411541           121        0.26 

OBSERVATION

• The above plot shows us the Top-25 Operating Physicians with most percentage of Fraudulent Case Submissions.
  • Here, PHY330576 has the highest percentage share of fraudulent cases. But, the difference b/w other providers is not so high.

Q6. Which are the Top-25 Operating Physicians with maximum number of non-fraudulent cases?

tmp_only_non_frauds = tmp[tmp['Fraud?'] == 'No'].sort_values(by=['Percentage'], ascending=False).reset_index(drop=True)

print(tmp_only_non_frauds[['OperatingPhysician','Num_of_cases','Percentage']].head(25), "\n")

with plt.style.context('seaborn'):
    plt.figure(figsize=(14,8))
    fig = sns.barplot(data=tmp_only_non_frauds.iloc[0:25], x="OperatingPhysician", y="Num_of_cases", palette='Accent')
    # Using the "patches" function we will get the location of the rectangle bars from the graph.
    ## Then by using those location(width & height) values we will add the annotations
    for p in fig.patches:
        width = p.get_width()
        height = p.get_height()
        x, y = p.get_xy()
        fig.annotate(f'{str(round((height*100)/tot_non_fraud_cases,2))+"%"}', (x + width/2, y + height*1.025), ha='center', fontsize=13.5, rotation=90)
    
    # Providing the labels and title to the graph
    plt.xlabel("\nTop Non-Fraudulent OperatingPhysician", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Number (or % share) of Cases\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Top-25 OperatingPhysician with most number of non-fraudulent cases\n", fontdict=title_font_dict)
    plt.plot();

   OperatingPhysician  Num_of_cases  Percentage
0           PHY387900           180        0.27
1           PHY351121           179        0.27
2           PHY375943           147        0.22
3           PHY367255           132        0.20
4           PHY432614           129        0.19
5           PHY362889           122        0.18
6           PHY326984           121        0.18
7           PHY356444           108        0.16
8           PHY360179            98        0.15
9           PHY405720            96        0.14
10          PHY373032            97        0.14
11          PHY321493            88        0.13
12          PHY342223            89        0.13
13          PHY318242            85        0.13
14          PHY319973            89        0.13
15          PHY326049            81        0.12
16          PHY366914            79        0.12
17          PHY361063            84        0.12
18          PHY388040            82        0.12
19          PHY322775            71        0.11
20          PHY313705            77        0.11
21          PHY403755            77        0.11
22          PHY351973            69        0.10
23          PHY405650            69        0.10
24          PHY415621            68        0.10 

OBSERVATION

• The above plot shows us the Top-25 Operating Physicians with most percentage of Non-Fraudulent Case Submissions.
  • Here, PHY387900 and PHY351121 has the highest percentage share of non-fraudulent cases. But, the difference b/w providers is not so high.

Q7. Which are the Top-25 Other Physicians with maximum number of fraudulent cases?

tmp = pd.DataFrame(train_iobp_df.groupby(['OtherPhysician','PotentialFraud'])['BeneID'].count()).reset_index()
tmp.columns = ['OtherPhysician', 'Fraud?', 'Num_of_cases']
tot_fraud_cases = tmp[tmp['Fraud?'] == 'Yes']['Num_of_cases'].sum()
tot_non_fraud_cases = tmp[tmp['Fraud?'] == 'No']['Num_of_cases'].sum()
tmp['Cases'] = tmp['Fraud?'].apply(lambda val: tot_non_fraud_cases if val == "No" else tot_fraud_cases)
tmp['Percentage'] = round(((tmp['Num_of_cases'] / tmp['Cases']) * 100),2)

tmp.head()

	OtherPhysician	Fraud?	Num_of_cases	Cases	Percentage
0	PHY311001	No	1	125093	0.0
1	PHY311003	Yes	2	74643	0.0
2	PHY311005	No	2	125093	0.0
3	PHY311006	No	3	125093	0.0
4	PHY311007	Yes	1	74643	0.0

tmp_only_frauds = tmp[tmp['Fraud?'] == 'Yes'].sort_values(by=['Percentage'], ascending=False).reset_index(drop=True)

print(tmp_only_frauds[['OtherPhysician','Num_of_cases','Percentage']].head(25), "\n")

with plt.style.context('seaborn'):
    plt.figure(figsize=(14,8))
    fig = sns.barplot(data=tmp_only_frauds.iloc[0:25], x="OtherPhysician", y="Num_of_cases", palette='Accent')
    # Using the "patches" function we will get the location of the rectangle bars from the graph.
    ## Then by using those location(width & height) values we will add the annotations
    for p in fig.patches:
        width = p.get_width()
        height = p.get_height()
        x, y = p.get_xy()
        fig.annotate(f'{str(round((height*100)/tot_fraud_cases,2))+"%"}', (x + width/2, y + height*1.025), ha='center', fontsize=13.5, rotation=90)
    
    # Providing the labels and title to the graph
    plt.xlabel("\nTop Fraudulent OtherPhysician", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Number (or % share) of Cases\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Top-25 OtherPhysician with most number of fraudulent cases\n", fontdict=title_font_dict)
    plt.plot();

   OtherPhysician  Num_of_cases  Percentage
0       PHY412132          1247        1.67
1       PHY341578          1098        1.47
2       PHY338032          1070        1.43
3       PHY337425          1041        1.39
4       PHY347064           806        1.08
5       PHY322092           771        1.03
6       PHY409965           744        1.00
7       PHY313818           730        0.98
8       PHY350277           682        0.91
9       PHY415321           678        0.91
10      PHY313278           625        0.84
11      PHY359122           614        0.82
12      PHY416093           538        0.72
13      PHY333735           496        0.66
14      PHY421058           432        0.58
15      PHY359931           414        0.55
16      PHY396637           400        0.54
17      PHY327964           371        0.50
18      PHY336024           372        0.50
19      PHY315344           363        0.49
20      PHY410597           357        0.48
21      PHY416732           356        0.48
22      PHY383336           344        0.46
23      PHY325906           293        0.39
24      PHY356051           283        0.38 

OBSERVATION

• The above plot shows us the Top-25 Other Physicians with most percentage of Fraudulent Case Submissions.
  • Here, PHY412132 has the highest percentage share of fraudulent cases. But, the difference with other providers is not so high.

Q8. Which are the Top-25 Other Physicians with maximum number of non-fraudulent cases?

tmp_only_non_frauds = tmp[tmp['Fraud?'] == 'No'].sort_values(by=['Percentage'], ascending=False).reset_index(drop=True)

print(tmp_only_non_frauds[['OtherPhysician','Num_of_cases','Percentage']].head(25), "\n")

with plt.style.context('seaborn'):
    plt.figure(figsize=(14,8))
    fig = sns.barplot(data=tmp_only_non_frauds.iloc[0:25], x="OtherPhysician", y="Num_of_cases", palette='Accent')
    # Using the "patches" function we will get the location of the rectangle bars from the graph.
    ## Then by using those location(width & height) values we will add the annotations
    for p in fig.patches:
        width = p.get_width()
        height = p.get_height()
        x, y = p.get_xy()
        fig.annotate(f'{str(round((height*100)/tot_non_fraud_cases,2))+"%"}', (x + width/2, y + height*1.025), ha='center', fontsize=13.5, rotation=90)
    
    # Providing the labels and title to the graph
    plt.xlabel("\nTop Non-Fraudulent OtherPhysician", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Number (or % share) of Cases\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Top-25 OtherPhysician with most number of non-fraudulent cases\n", fontdict=title_font_dict)
    plt.plot();

   OtherPhysician  Num_of_cases  Percentage
0       PHY422235           369        0.29
1       PHY387900           351        0.28
2       PHY375943           328        0.26
3       PHY411722           313        0.25
4       PHY363309           305        0.24
5       PHY331484           295        0.24
6       PHY362889           294        0.24
7       PHY326984           292        0.23
8       PHY367255           269        0.22
9       PHY389456           271        0.22
10      PHY356444           258        0.21
11      PHY382485           249        0.20
12      PHY315424           243        0.19
13      PHY405720           220        0.18
14      PHY354358           217        0.17
15      PHY323501           202        0.16
16      PHY403755           196        0.16
17      PHY384394           196        0.16
18      PHY388477           197        0.16
19      PHY395822           195        0.16
20      PHY326286           189        0.15
21      PHY415621           182        0.15
22      PHY411662           189        0.15
23      PHY359164           182        0.15
24      PHY342223           184        0.15 

OBSERVATION

• The above plot shows us the Top-25 Other Physicians with most percentage of Non-Fraudulent Case Submissions.
  • Here, PHY422235 has the highest percentage share of non-fraudulent cases. But, the difference with other providers is not so high.

Q9. Which are the Top-25 ClmAdmitDiagnosisCode with maximum number of fraudulent cases?

tmp = pd.DataFrame(train_iobp_df.groupby(['ClmAdmitDiagnosisCode','PotentialFraud'])['BeneID'].count()).reset_index()
tmp.columns = ['ClmAdmitDiagnosisCode', 'Fraud?', 'Num_of_cases']
tot_fraud_cases = tmp[tmp['Fraud?'] == 'Yes']['Num_of_cases'].sum()
tot_non_fraud_cases = tmp[tmp['Fraud?'] == 'No']['Num_of_cases'].sum()
tmp['Cases'] = tmp['Fraud?'].apply(lambda val: tot_non_fraud_cases if val == "No" else tot_fraud_cases)
tmp['Percentage'] = round(((tmp['Num_of_cases'] / tmp['Cases']) * 100),2)

tmp.head()

	ClmAdmitDiagnosisCode	Fraud?	Num_of_cases	Cases	Percentage
0	0030	No	1	83850	0.00
1	0030	Yes	1	62049	0.00
2	0059	No	1	83850	0.00
3	0059	Yes	2	62049	0.00
4	00845	No	31	83850	0.04

tmp_only_frauds = tmp[tmp['Fraud?'] == 'Yes'].sort_values(by=['Percentage'], ascending=False).reset_index(drop=True)

print(tmp_only_frauds[['ClmAdmitDiagnosisCode','Num_of_cases','Percentage']].head(25), "\n")

with plt.style.context('seaborn'):
    plt.figure(figsize=(14,8))
    fig = sns.barplot(data=tmp_only_frauds.iloc[0:25], x="ClmAdmitDiagnosisCode", y="Num_of_cases", palette='Accent')
    # Using the "patches" function we will get the location of the rectangle bars from the graph.
    ## Then by using those location(width & height) values we will add the annotations
    for p in fig.patches:
        width = p.get_width()
        height = p.get_height()
        x, y = p.get_xy()
        fig.annotate(f'{str(round((height*100)/tot_fraud_cases,2))+"%"}', (x + width/2, y + height*1.025), ha='center', fontsize=13.5, rotation=90)
    
    # Providing the labels and title to the graph
    plt.xlabel("\nTop Fraudulent ClmAdmitDiagnosisCode", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Number (or % share) of Cases\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Top-25 ClmAdmitDiagnosisCode with most number of fraudulent cases\n", fontdict=title_font_dict)
    plt.plot();

   ClmAdmitDiagnosisCode  Num_of_cases  Percentage
0                  42731          1529        2.46
1                  V7612          1441        2.32
2                  78605          1432        2.31
3                  78650          1191        1.92
4                  78900          1020        1.64
5                   4019          1006        1.62
6                  25000           873        1.41
7                    486           843        1.36
8                  78079           828        1.33
9                   7802           793        1.28
10                  7295           740        1.19
11                  5990           703        1.13
12                 V5883           691        1.11
13                  4280           661        1.07
14                  7242           600        0.97
15                  7862           593        0.96
16                 V5789           573        0.92
17                 V5861           556        0.90
18                  2724           549        0.88
19                  V571           538        0.87
20                 78097           536        0.86
21                 78609           521        0.84
22                 41401           507        0.82
23                 78659           435        0.70
24                  7804           417        0.67 

OBSERVATION

• The above plot shows us the Top-25 'Claim Admit Diagnosis Code' with most percentage of Fraudulent Case Submissions.

Q10. Which are the Top-25 ClmAdmitDiagnosisCode with maximum number of non-fraudulent cases?

tmp_only_non_frauds = tmp[tmp['Fraud?'] == 'No'].sort_values(by=['Percentage'], ascending=False).reset_index(drop=True)

print(tmp_only_non_frauds[['ClmAdmitDiagnosisCode','Num_of_cases','Percentage']].head(25), "\n")

with plt.style.context('seaborn'):
    plt.figure(figsize=(14,8))
    fig = sns.barplot(data=tmp_only_non_frauds.iloc[0:25], x="ClmAdmitDiagnosisCode", y="Num_of_cases", palette='Accent')
    # Using the "patches" function we will get the location of the rectangle bars from the graph.
    ## Then by using those location(width & height) values we will add the annotations
    for p in fig.patches:
        width = p.get_width()
        height = p.get_height()
        x, y = p.get_xy()
        fig.annotate(f'{str(round((height*100)/tot_non_fraud_cases,2))+"%"}', (x + width/2, y + height*1.025), ha='center', fontsize=13.5, rotation=90)
    
    # Providing the labels and title to the graph
    plt.xlabel("\nTop Non-Fraudulent ClmAdmitDiagnosisCode", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Number (or % share) of Cases\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Top-25 ClmAdmitDiagnosisCode with most number of non-fraudulent cases\n", fontdict=title_font_dict)
    plt.plot();

   ClmAdmitDiagnosisCode  Num_of_cases  Percentage
0                  V7612          2633        3.14
1                  42731          2105        2.51
2                   4019          1726        2.06
3                  78605          1560        1.86
4                  25000          1495        1.78
5                  78900          1316        1.57
6                  V5883          1182        1.41
7                   7295          1104        1.32
8                  78650          1082        1.29
9                   7242           997        1.19
10                 V5861           980        1.17
11                  2724           958        1.14
12                 78079           951        1.13
13                  5990           922        1.10
14                  V571           905        1.08
15                  7862           868        1.04
16                  7802           820        0.98
17                  4011           660        0.79
18                   486           621        0.74
19                  7804           622        0.74
20                 41401           613        0.73
21                  7245           596        0.71
22                 78609           598        0.71
23                 71946           573        0.68
24                  7840           551        0.66 

OBSERVATION

• The above plot shows us the Top-25 'Claim Admit Diagnosis Code' with most percentage of Non-Fraudulent Case Submissions.

• Main observation from the above 2 plots is that same Claim Admit Diagnostic Codes have similar percentages for false and no-false claims. Therefore, it feels like this feature might not be very useful.

Q11. Which are the Top-25 DiagnosisGroupCode with maximum number of fraudulent cases?

tmp = pd.DataFrame(train_iobp_df.groupby(['DiagnosisGroupCode','PotentialFraud'])['BeneID'].count()).reset_index()
tmp.columns = ['DiagnosisGroupCode', 'Fraud?', 'Num_of_cases']
tot_fraud_cases = tmp[tmp['Fraud?'] == 'Yes']['Num_of_cases'].sum()
tot_non_fraud_cases = tmp[tmp['Fraud?'] == 'No']['Num_of_cases'].sum()
tmp['Cases'] = tmp['Fraud?'].apply(lambda val: tot_non_fraud_cases if val == "No" else tot_fraud_cases)
tmp['Percentage'] = round(((tmp['Num_of_cases'] / tmp['Cases']) * 100),2)

tmp.head()

	DiagnosisGroupCode	Fraud?	Num_of_cases	Cases	Percentage
0	000	No	57	17072	0.33
1	000	Yes	77	23402	0.33
2	001	No	2	17072	0.01
3	001	Yes	8	23402	0.03
4	002	No	9	17072	0.05

tmp_only_frauds = tmp[tmp['Fraud?'] == 'Yes'].sort_values(by=['Percentage'], ascending=False).reset_index(drop=True)

print(tmp_only_frauds[['DiagnosisGroupCode','Num_of_cases','Percentage']].head(25), "\n")

with plt.style.context('seaborn'):
    plt.figure(figsize=(14,8))
    fig = sns.barplot(data=tmp_only_frauds.iloc[0:25], x="DiagnosisGroupCode", y="Num_of_cases", palette='Accent')
    # Using the "patches" function we will get the location of the rectangle bars from the graph.
    ## Then by using those location(width & height) values we will add the annotations
    for p in fig.patches:
        width = p.get_width()
        height = p.get_height()
        x, y = p.get_xy()
        fig.annotate(f'{str(round((height*100)/tot_fraud_cases,2))+"%"}', (x + width/2, y + height*1.025), ha='center', fontsize=13.5, rotation=90)
    
    # Providing the labels and title to the graph
    plt.xlabel("\nTop Fraudulent DiagnosisGroupCode", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Number (or % share) of Cases\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Top-25 DiagnosisGroupCode with most number of fraudulent cases\n", fontdict=title_font_dict)
    plt.plot();

   DiagnosisGroupCode  Num_of_cases  Percentage
0                 882           111        0.47
1                 166           107        0.46
2                 186           102        0.44
3                 192            98        0.42
4                 945            95        0.41
5                 884            97        0.41
6                 939            96        0.41
7                 202            95        0.41
8                 883            96        0.41
9                 188            97        0.41
10                168            94        0.40
11                949            94        0.40
12                204            94        0.40
13                885            92        0.39
14                876            90        0.38
15                196            89        0.38
16                198            89        0.38
17                183            89        0.38
18                887            90        0.38
19                950            90        0.38
20                177            86        0.37
21                946            87        0.37
22                947            86        0.37
23                164            87        0.37
24                184            84        0.36 

OBSERVATION

• The above plot shows us the Top-25 'Diagnosis Group Code' with most percentage of Fraudulent Case Submissions.

Q12. Which are the Top-25 DiagnosisGroupCode with maximum number of non-fraudulent cases?

tmp_only_non_frauds = tmp[tmp['Fraud?'] == 'No'].sort_values(by=['Percentage'], ascending=False).reset_index(drop=True)

print(tmp_only_non_frauds[['DiagnosisGroupCode','Num_of_cases','Percentage']].head(25), "\n")

with plt.style.context('seaborn'):
    plt.figure(figsize=(14,8))
    fig = sns.barplot(data=tmp_only_non_frauds.iloc[0:25], x="DiagnosisGroupCode", y="Num_of_cases", palette='Accent')
    # Using the "patches" function we will get the location of the rectangle bars from the graph.
    ## Then by using those location(width & height) values we will add the annotations
    for p in fig.patches:
        width = p.get_width()
        height = p.get_height()
        x, y = p.get_xy()
        fig.annotate(f'{str(round((height*100)/tot_non_fraud_cases,2))+"%"}', (x + width/2, y + height*1.025), ha='center', fontsize=13.5, rotation=90)
    
    # Providing the labels and title to the graph
    plt.xlabel("\nTop Non-Fraudulent DiagnosisGroupCode", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Number (or % share) of Cases\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Top-25 DiagnosisGroupCode with most number of non-fraudulent cases\n", fontdict=title_font_dict)
    plt.plot();

   DiagnosisGroupCode  Num_of_cases  Percentage
0                 183            76        0.45
1                 167            76        0.45
2                 884            77        0.45
3                 208            75        0.44
4                 940            70        0.41
5                 881            70        0.41
6                 882            68        0.40
7                 187            69        0.40
8                 887            69        0.40
9                 941            69        0.40
10                939            67        0.39
11                876            66        0.39
12                168            66        0.39
13                182            67        0.39
14                184            65        0.38
15                205            65        0.38
16                206            65        0.38
17                177            64        0.37
18                190            63        0.37
19                283            64        0.37
20                198            63        0.37
21                204            64        0.37
22                180            64        0.37
23                948            64        0.37
24                880            63        0.37 

OBSERVATION

• The above plot shows us the Top-25 'Diagnosis Group Code' with most percentage of Non-Fraudulent Case Submissions.

• Main observation from the above 2 plots is that same Diagnosis Group Codes have similar percentages for false and no-false claims. Therefore, it feels like this feature might not be very useful.

Q13. Does Age_groups have any relationship with maximum number of fraudulent cases?

def bene_age_brackets(val):
    """
    Description : This function is created for allocating the age groups based on Beneficiary Age.
    """
    if val >=1 and val <=40:
        return 'Young'
    elif val > 40 and val <=60:
        return 'Mid'
    elif val > 60 and val <= 80:
        return 'Old'
    else:
        return 'Very Old'

train_iobp_df['AGE_groups'] = train_iobp_df['Bene_Age'].apply(lambda age: bene_age_brackets(age))

tmp = pd.DataFrame(train_iobp_df.groupby(['AGE_groups','PotentialFraud'])['BeneID'].count()).reset_index()
tmp.columns = ['AGE_groups', 'Fraud?', 'Num_of_cases']
tot_fraud_cases = tmp[tmp['Fraud?'] == 'Yes']['Num_of_cases'].sum()
tot_non_fraud_cases = tmp[tmp['Fraud?'] == 'No']['Num_of_cases'].sum()
tmp['Cases'] = tmp['Fraud?'].apply(lambda val: tot_non_fraud_cases if val == "No" else tot_fraud_cases)
tmp['Percentage'] = round(((tmp['Num_of_cases'] / tmp['Cases']) * 100),2)

tmp.head()

	AGE_groups	Fraud?	Num_of_cases	Cases	Percentage
0	Mid	No	35524	345415	10.28
1	Mid	Yes	21152	212796	9.94
2	Old	No	190334	345415	55.10
3	Old	Yes	116676	212796	54.83
4	Very Old	No	110885	345415	32.10

tmp_only_frauds = tmp[tmp['Fraud?'] == 'Yes'].sort_values(by=['Percentage'], ascending=False).reset_index(drop=True)

print(tmp_only_frauds[['AGE_groups','Num_of_cases','Percentage']].head(25), "\n")

with plt.style.context('seaborn'):
    plt.figure(figsize=(10,8))
    fig = sns.barplot(data=tmp_only_frauds, x="AGE_groups", y="Num_of_cases", palette='Accent')
    # Using the "patches" function we will get the location of the rectangle bars from the graph.
    ## Then by using those location(width & height) values we will add the annotations
    for p in fig.patches:
        width = p.get_width()
        height = p.get_height()
        x, y = p.get_xy()
        fig.annotate(f'{str(round((height*100)/tot_fraud_cases,2))+"%"}', (x + width/2, y + height*1.025), ha='center', fontsize=13.5, rotation=0)
    
    # Providing the labels and title to the graph
    plt.xlabel("\nTop Fraudulent AGE_groups", fontdict=label_font_dict)
    plt.xticks(rotation=0, fontsize=12)
    plt.ylabel("Number (or % share) of Cases\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("AGE_groups with most number of fraudulent cases\n", fontdict=title_font_dict)
    plt.plot();

  AGE_groups  Num_of_cases  Percentage
0        Old        116676       54.83
1   Very Old         69780       32.79
2        Mid         21152        9.94
3      Young          5188        2.44 

OBSERVATION

• The above plot shows us the percentage of Fraudulent Case Submissions for various Age Groups.

Q14. Does Age_groups have any relationship with maximum number of non-fraudulent cases?

tmp_only_non_frauds = tmp[tmp['Fraud?'] == 'No'].sort_values(by=['Percentage'], ascending=False).reset_index(drop=True)

print(tmp_only_non_frauds[['AGE_groups','Num_of_cases','Percentage']].head(25), "\n")

with plt.style.context('seaborn'):
    plt.figure(figsize=(10,8))
    fig = sns.barplot(data=tmp_only_non_frauds, x="AGE_groups", y="Num_of_cases", palette='Accent')
    # Using the "patches" function we will get the location of the rectangle bars from the graph.
    ## Then by using those location(width & height) values we will add the annotations
    for p in fig.patches:
        width = p.get_width()
        height = p.get_height()
        x, y = p.get_xy()
        fig.annotate(f'{str(round((height*100)/tot_non_fraud_cases,2))+"%"}', (x + width/2, y + height*1.025), ha='center', fontsize=13.5, rotation=0)
    
    # Providing the labels and title to the graph
    plt.xlabel("\nTop Non-Fraudulent AGE_groups", fontdict=label_font_dict)
    plt.xticks(rotation=0, fontsize=12)
    plt.ylabel("Number (or % share) of Cases\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("AGE_groups with most number of non-fraudulent cases\n", fontdict=title_font_dict)
    plt.plot();

  AGE_groups  Num_of_cases  Percentage
0        Old        190334       55.10
1   Very Old        110885       32.10
2        Mid         35524       10.28
3      Young          8672        2.51 

OBSERVATION

• The above plot shows us the percentage of Non-Fraudulent Case Submissions for various Age Groups.

• Main observation from the above 2 plots is that same Age Groups have similar percentages for false and no-false claims. Therefore, it feels like this feature might not be very useful.

Q15. Which are the Top-25 States with maximum number of fraudulent cases?

tmp = pd.DataFrame(train_iobp_df.groupby(['State','PotentialFraud'])['BeneID'].count()).reset_index()
tmp.columns = ['State', 'Fraud?', 'Num_of_cases']
tot_fraud_cases = tmp[tmp['Fraud?'] == 'Yes']['Num_of_cases'].sum()
tot_non_fraud_cases = tmp[tmp['Fraud?'] == 'No']['Num_of_cases'].sum()
tmp['Cases'] = tmp['Fraud?'].apply(lambda val: tot_non_fraud_cases if val == "No" else tot_fraud_cases)
tmp['Percentage'] = round(((tmp['Num_of_cases'] / tmp['Cases']) * 100),2)

tmp.head()

	State	Fraud?	Num_of_cases	Cases	Percentage
0	1	No	6715	345415	1.94
1	1	Yes	3525	212796	1.66
2	2	No	531	345415	0.15
3	2	Yes	207	212796	0.10
4	3	No	7314	345415	2.12

tmp_only_frauds = tmp[tmp['Fraud?'] == 'Yes'].sort_values(by=['Percentage'], ascending=False).reset_index(drop=True)

print(tmp_only_frauds[['State','Num_of_cases','Percentage']].head(25), "\n")

with plt.style.context('seaborn'):
    plt.figure(figsize=(14,8))
    fig = sns.barplot(data=tmp_only_frauds.iloc[0:25], x="State", y="Num_of_cases", palette='Accent')
    # Using the "patches" function we will get the location of the rectangle bars from the graph.
    ## Then by using those location(width & height) values we will add the annotations
    for p in fig.patches:
        width = p.get_width()
        height = p.get_height()
        x, y = p.get_xy()
        fig.annotate(f'{str(round((height*100)/tot_fraud_cases,2))+"%"}', (x + width/2, y + height*1.025), ha='center', fontsize=13.5, rotation=90)
    
    # Providing the labels and title to the graph
    plt.xlabel("\nTop Fraudulent State Codes", fontdict=label_font_dict)
    plt.xticks(rotation=0, fontsize=12)
    plt.ylabel("Number (or % share) of Cases\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Top-25 State Codes with most number of fraudulent cases\n", fontdict=title_font_dict)
    plt.plot();

    State  Num_of_cases  Percentage
0       5         30335       14.26
1      10         17512        8.23
2      33         17492        8.22
3      39         11448        5.38
4      45         10135        4.76
5      31          9112        4.28
6      49          8613        4.05
7      23          8538        4.01
8      14          8509        4.00
9      22          7798        3.66
10     44          6709        3.15
11     36          6381        3.00
12     26          5301        2.49
13     50          4782        2.25
14     15          4635        2.18
15     34          4385        2.06
16     11          4123        1.94
17      6          3666        1.72
18      1          3525        1.66
19     24          3453        1.62
20     42          3180        1.49
21     16          2733        1.28
22     21          2576        1.21
23     46          2124        1.00
24      3          2030        0.95 

OBSERVATION

• The above plot shows us the Top-25 State Codes with most percentage of Fraudulent Case Submissions.

Q16. What are the Top-25 States with maximum number of non-fraudulent cases?

tmp_only_non_frauds = tmp[tmp['Fraud?'] == 'No'].sort_values(by=['Percentage'], ascending=False).reset_index(drop=True)

print(tmp_only_non_frauds[['State','Num_of_cases','Percentage']].head(25), "\n")

with plt.style.context('seaborn'):
    plt.figure(figsize=(14,8))
    fig = sns.barplot(data=tmp_only_non_frauds.iloc[0:25], x="State", y="Num_of_cases", palette='Accent')
    # Using the "patches" function we will get the location of the rectangle bars from the graph.
    ## Then by using those location(width & height) values we will add the annotations
    for p in fig.patches:
        width = p.get_width()
        height = p.get_height()
        x, y = p.get_xy()
        fig.annotate(f'{str(round((height*100)/tot_non_fraud_cases,2))+"%"}', (x + width/2, y + height*1.025), ha='center', fontsize=13.5, rotation=90)
    
    # Providing the labels and title to the graph
    plt.xlabel("\nTop Non-Fraudulent State Codes", fontdict=label_font_dict)
    plt.xticks(rotation=0, fontsize=12)
    plt.ylabel("Number (or % share) of Cases\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Top-25 State Codes with most number of non-fraudulent cases\n", fontdict=title_font_dict)
    plt.plot();

    State  Num_of_cases  Percentage
0      45         23887        6.92
1      10         21561        6.24
2       5         21015        6.08
3      33         17532        5.08
4      14         15908        4.61
5      36         14910        4.32
6      34         14520        4.20
7      11         12880        3.73
8      39         12803        3.71
9      23         12805        3.71
10     15          9578        2.77
11     21          8685        2.51
12     18          8643        2.50
13     52          7839        2.27
14     44          7709        2.23
15     26          7610        2.20
16     42          7311        2.12
17      3          7314        2.12
18     50          6958        2.01
19     31          6828        1.98
20      1          6715        1.94
21     49          6384        1.85
22     19          6230        1.80
23     25          5863        1.70
24     22          5826        1.69 

OBSERVATION

• The above plot shows us the Top-25 State Codes with most percentage of Non-Fraudulent Case Submissions.

• Main observation from the above 2 plots is that same State Codes have similar percentages for false and no-false claims. Therefore, it feels like this feature might not be very useful.

Q17. Which are the Top-25 Country with maximum number of fraudulent cases?

tmp = pd.DataFrame(train_iobp_df.groupby(['County','PotentialFraud'])['BeneID'].count()).reset_index()
tmp.columns = ['County', 'Fraud?', 'Num_of_cases']
tot_fraud_cases = tmp[tmp['Fraud?'] == 'Yes']['Num_of_cases'].sum()
tot_non_fraud_cases = tmp[tmp['Fraud?'] == 'No']['Num_of_cases'].sum()
tmp['Cases'] = tmp['Fraud?'].apply(lambda val: tot_non_fraud_cases if val == "No" else tot_fraud_cases)
tmp['Percentage'] = round(((tmp['Num_of_cases'] / tmp['Cases']) * 100),2)

tmp.head()

	County	Fraud?	Num_of_cases	Cases	Percentage
0	0	No	6584	345415	1.91
1	0	Yes	4897	212796	2.30
2	1	No	8	345415	0.00
3	1	Yes	4	212796	0.00
4	10	No	10750	345415	3.11

tmp_only_frauds = tmp[tmp['Fraud?'] == 'Yes'].sort_values(by=['Percentage'], ascending=False).reset_index(drop=True)

print(tmp_only_frauds[['County','Num_of_cases','Percentage']].head(25), "\n")

with plt.style.context('seaborn'):
    plt.figure(figsize=(15,10))
    fig = sns.barplot(data=tmp_only_frauds.iloc[0:25], x="County", y="Num_of_cases", palette='Accent')
    # Using the "patches" function we will get the location of the rectangle bars from the graph.
    ## Then by using those location(width & height) values we will add the annotations
    for p in fig.patches:
        width = p.get_width()
        height = p.get_height()
        x, y = p.get_xy()
        fig.annotate(f'{str(round((height*100)/tot_fraud_cases,2))+"%"}', (x + width/2, y + height*1.025), ha='center', fontsize=13.5, rotation=90)
    
    # Providing the labels and title to the graph
    plt.xlabel("\nTop Fraudulent Country Codes", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Number (or % share) of Cases\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Top-25 Country Codes with most number of fraudulent cases\n", fontdict=title_font_dict)
    plt.plot();

    County  Num_of_cases  Percentage
0      200         10078        4.74
1      470          7048        3.31
2      400          5962        2.80
3      590          5814        2.73
4        0          4897        2.30
5      160          4865        2.29
6      620          4704        2.21
7      130          4551        2.14
8      490          4477        2.10
9      170          4362        2.05
10     440          4314        2.03
11      20          4174        1.96
12      90          4079        1.92
13     150          3844        1.81
14     290          3725        1.75
15     510          3505        1.65
16     310          3437        1.62
17     390          3387        1.59
18     331          3284        1.54
19      10          3232        1.52
20     141          3202        1.50
21      60          3181        1.49
22     700          3020        1.42
23     250          2999        1.41
24     530          2871        1.35 

OBSERVATION

• The above plot shows us the Top-25 Country Codes with most percentage of Fraudulent Case Submissions.

Q18. What are the Top-25 Country with maximum number of non-fraudulent cases?

tmp_only_non_frauds = tmp[tmp['Fraud?'] == 'No'].sort_values(by=['Percentage'], ascending=False).reset_index(drop=True)

print(tmp_only_non_frauds[['County','Num_of_cases','Percentage']].head(25), "\n")

with plt.style.context('seaborn'):
    plt.figure(figsize=(15,10))
    fig = sns.barplot(data=tmp_only_non_frauds.iloc[0:25], x="County", y="Num_of_cases", palette='Accent')
    # Using the "patches" function we will get the location of the rectangle bars from the graph.
    ## Then by using those location(width & height) values we will add the annotations
    for p in fig.patches:
        width = p.get_width()
        height = p.get_height()
        x, y = p.get_xy()
        fig.annotate(f'{str(round((height*100)/tot_non_fraud_cases,2))+"%"}', (x + width/2, y + height*1.025), ha='center', fontsize=13.5, rotation=90)
    
    # Providing the labels and title to the graph
    plt.xlabel("\nTop Non-Fraudulent Country Codes", fontdict=label_font_dict)
    plt.xticks(rotation=90, fontsize=12)
    plt.ylabel("Number (or % share) of Cases\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Top-25 Country Codes with most number of non-fraudulent cases\n", fontdict=title_font_dict)
    plt.plot();

    County  Num_of_cases  Percentage
0       10         10750        3.11
1       60          8814        2.55
2       20          8458        2.45
3       90          7007        2.03
4        0          6584        1.91
5      200          5879        1.70
6      150          5843        1.69
7      141          5793        1.68
8      400          5735        1.66
9      160          5668        1.64
10     310          5590        1.62
11      70          5487        1.59
12      50          5430        1.57
13     250          5385        1.56
14      40          5276        1.53
15     470          5230        1.51
16     480          5229        1.51
17     490          5010        1.45
18     100          4995        1.45
19     120          4936        1.43
20     240          4887        1.41
21     550          4513        1.31
22      30          4491        1.30
23     290          4417        1.28
24     390          4428        1.28 

OBSERVATION

• The above plot shows us the Top-25 Country Codes with most percentage of Non-Fraudulent Case Submissions.

• Main observation from the above 2 plots is that same Country Codes have similar percentages for false and no-false claims. Therefore, it feels like this feature might not be very useful.

Q19. Does various Human Races have any relationship with maximum number of fraudulent cases?

tmp = pd.DataFrame(train_iobp_df.groupby(['Race','PotentialFraud'])['BeneID'].count()).reset_index()
tmp.columns = ['Race', 'Fraud?', 'Num_of_cases']
tot_fraud_cases = tmp[tmp['Fraud?'] == 'Yes']['Num_of_cases'].sum()
tot_non_fraud_cases = tmp[tmp['Fraud?'] == 'No']['Num_of_cases'].sum()
tmp['Cases'] = tmp['Fraud?'].apply(lambda val: tot_non_fraud_cases if val == "No" else tot_fraud_cases)
tmp['Percentage'] = round(((tmp['Num_of_cases'] / tmp['Cases']) * 100),2)

tmp.head()

	Race	Fraud?	Num_of_cases	Cases	Percentage
0	1	No	292691	345415	84.74
1	1	Yes	178345	212796	83.81
2	2	No	35356	345415	10.24
3	2	Yes	20284	212796	9.53
4	3	No	10753	345415	3.11

tmp_only_frauds = tmp[tmp['Fraud?'] == 'Yes'].sort_values(by=['Percentage'], ascending=False).reset_index(drop=True)

print(tmp_only_frauds[['Race','Num_of_cases','Percentage']].head(25), "\n")

with plt.style.context('seaborn'):
    plt.figure(figsize=(10,8))
    fig = sns.barplot(data=tmp_only_frauds, x="Race", y="Num_of_cases", palette='Accent')
    # Using the "patches" function we will get the location of the rectangle bars from the graph.
    ## Then by using those location(width & height) values we will add the annotations
    for p in fig.patches:
        width = p.get_width()
        height = p.get_height()
        x, y = p.get_xy()
        fig.annotate(f'{str(round((height*100)/tot_fraud_cases,2))+"%"}', (x + width/2, y + height*1.025), ha='center', fontsize=13.5, rotation=0)
    
    # Providing the labels and title to the graph
    plt.xlabel("\nTop Fraudulent Race", fontdict=label_font_dict)
    plt.xticks(rotation=0, fontsize=12)
    plt.ylabel("Number (or % share) of Cases\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Human Race with most number of fraudulent cases\n", fontdict=title_font_dict)
    plt.plot();

   Race  Num_of_cases  Percentage
0     1        178345       83.81
1     2         20284        9.53
2     3          8962        4.21
3     5          5205        2.45 

OBSERVATION

• The above plot shows us the percentage of Fraudulent Case Submissions for various Human Races.

Q20. Does various Human Races have any relationship with maximum number of non-fraudulent cases?

tmp_only_non_frauds = tmp[tmp['Fraud?'] == 'No'].sort_values(by=['Percentage'], ascending=False).reset_index(drop=True)

print(tmp_only_non_frauds[['Race','Num_of_cases','Percentage']].head(25), "\n")

with plt.style.context('seaborn'):
    plt.figure(figsize=(10,8))
    fig = sns.barplot(data=tmp_only_non_frauds, x="Race", y="Num_of_cases", palette='Accent')
    # Using the "patches" function we will get the location of the rectangle bars from the graph.
    ## Then by using those location(width & height) values we will add the annotations
    for p in fig.patches:
        width = p.get_width()
        height = p.get_height()
        x, y = p.get_xy()
        fig.annotate(f'{str(round((height*100)/tot_non_fraud_cases,2))+"%"}', (x + width/2, y + height*1.025), ha='center', fontsize=13.5, rotation=0)
    
    # Providing the labels and title to the graph
    plt.xlabel("\nTop Non-Fraudulent Race", fontdict=label_font_dict)
    plt.xticks(rotation=0, fontsize=12)
    plt.ylabel("Number (or % share) of Cases\n", fontdict=label_font_dict)
    plt.minorticks_on()
    plt.grid(which='major', linestyle="--", color='lightgrey')
    plt.title("Human Race with most number of non-fraudulent cases\n", fontdict=title_font_dict)
    plt.plot();

   Race  Num_of_cases  Percentage
0     1        292691       84.74
1     2         35356       10.24
2     3         10753        3.11
3     5          6615        1.92 

OBSERVATION

• The above plot shows us the percentage of Non-Fraudulent Case Submissions for various Human Races.

• Main observation from the above 2 plots is that same Human Races have similar percentages for false and no-false claims. Therefore, it feels like this feature might not be very useful.